This study presents a direct prediction approach for estimating the prevalence of radiological findings from text, eliminating the need for computationally intensive similarity-based comparisons. The proposed method, named PRESTIGE, trades off apriori processing time in favor of real-time inference speed. PRESTIGE takes a single query example as input and estimates how many instances in a larger collection share similar characteristics. This approach is evaluated in both "global" settings (across an entire dataset) and "local" settings (within a relevant subset). In a clinical environment, the approach allows radiologists, hospital administrators, or researchers to quickly gauge the frequency of an observation occurring within their local patient population, without requiring manual chart reviews or text-mining. Experiments on radiology reports demonstrate that PRESTIGE achieves high accuracy in predicting prevalence, significantly reducing both inference time and storage requirements compared to naive or approximate nearest neighbor methods. Further efficiency gains are explored by incorporating low-rank adaptation (LoRA), a technique that updates only a small portion of the model's parameters, thereby reducing the model's size and accelerating training without introducing additional inference costs. Results indicate that LoRA can maintain strong performance while drastically lowering hardware demands, although a slight decrease in accuracy may occur depending on the dataset. Overall, these findings suggest that PRESTIGE can serve as an efficient and scalable solution for real-time, example-based prevalence estimation of radiological findings. By focusing on inference speed, this work facilitates interactive exploration of datasets and potentially streamlines research on determining how frequently specific observations appear in large document collections.

Size measurements of tumor manifestations on follow-up CT examinations are crucial for evaluating treatment outcomes in cancer patients. Efficient lesion segmentation can speed up these radiological workflows. While numerous benchmarks and challenges address lesion segmentation in specific organs like the liver, kidneys, and lungs, the larger variety of lesion types encountered in clinical practice demands a more universal approach. To address this gap, we introduced the ULS23 benchmark for 3D universal lesion segmentation in chest-abdomen-pelvis CT examinations. The ULS23 training dataset contains 38,693 lesions across this region, including challenging pancreatic, colon and bone lesions. For evaluation purposes, we curated a dataset comprising 775 lesions from 284 patients. Each of these lesions was identified as a target lesion in a clinical context, ensuring diversity and clinical relevance within this dataset. The ULS23 benchmark is publicly accessible at https://uls23.grand-challenge.org, enabling researchers worldwide to assess the performance of their segmentation methods. Furthermore, we have developed and publicly released our baseline semi-supervised 3D lesion segmentation model. This model achieved an average Dice coefficient of 0.703 ± 0.240 on the challenge test set. We invite ongoing submissions to advance the development of future ULS models.

Fetal growth restriction, affecting up to 10% of pregnancies, is a critical factor contributing to perinatal mortality and morbidity. Ultrasound measurements of the fetal abdominal circumference (AC) are a key aspect of monitoring fetal growth. However, the routine practice of biometric obstetric ultrasounds is limited in low-resource settings due to the high cost of sonography equipment and the scarcity of trained sonographers. To address this issue, we organized the ACOUSLIC-AI (Abdominal Circumference Operator-agnostic UltraSound measurement in Low-Income Countries) challenge to investigate the feasibility of automatically estimating fetal AC from blind-sweep ultrasound scans acquired by novice operators using low-cost devices. Training data, collected from three Public Health Units (PHUs) in Sierra Leone are made publicly available. Private validation and test sets, containing data from two PHUs in Tanzania and a European hospital, are provided through the Grand-Challenge platform. All sets were annotated by experienced readers. Sixteen international teams participated in this challenge, with six teams submitting to the Final Test Phase. In this article, we present the results of the three top-performing AI models from the ACOUSLIC-AI challenge, which are publicly accessible. We evaluate their performance in fetal abdomen frame selection, segmentation, abdominal circumference measurement, and compare their performance against clinical standards for fetal AC measurement. Clinical comparisons demonstrated that the limits of agreement (LoA) for A2 in fetal AC measurements are comparable to the interobserver LoA reported in the literature. The algorithms developed as part of the ACOUSLIC-AI challenge provide a benchmark for future algorithms on the selection and segmentation of fetal abdomen frames to further minimize fetal abdominal circumference measurement variability.

In recent years, several algorithms have been developed for the segmentation of the Inferior Alveolar Canal (IAC) in Cone-Beam Computed Tomography (CBCT) scans. However, the availability of public datasets in this domain is limited, resulting in a lack of comparative evaluation studies on a common benchmark. To address this scientific gap and encourage deep learning research in the field, the ToothFairy challenge was organized within the MICCAI 2023 conference. In this context, a public dataset was released to also serve as a benchmark for future research. The dataset comprises 443 CBCT scans, with voxel-level annotations of the IAC available for 153 of them, making it the largest publicly available dataset of its kind. The participants of the challenge were tasked with developing an algorithm to accurately identify the IAC using the 2D and 3D-annotated scans. This paper presents the details of the challenge and the contributions made by the most promising methods proposed by the participants. It represents the first comprehensive comparative evaluation of IAC segmentation methods on a common benchmark dataset, providing insights into the current state-of-the-art algorithms and outlining future research directions. Furthermore, to ensure reproducibility and promote future developments, an open-source repository that collects the implementations of the best submissions was released.

Challenges drive the state-of-the-art of automated medical image analysis. The quantity of public training data that they provide can limit the performance of their solutions. Public access to the training methodology for these solutions remains absent. This study implements the Type Three (T3) challenge format, which allows for training solutions on private data and guarantees reusable training methodologies. With T3, challenge organizers train a codebase provided by the participants on sequestered training data. T3 was implemented in the STOIC2021 challenge, with the goal of predicting from a computed tomography (CT) scan whether subjects had a severe COVID-19 infection, defined as intubation or death within one month. STOIC2021 consisted of a Qualification phase, where participants developed challenge solutions using 2000 publicly available CT scans, and a Final phase, where participants submitted their training methodologies with which solutions were trained on CT scans of 9724 subjects. The organizers successfully trained six of the eight Final phase submissions. The submitted codebases for training and running inference were released publicly. The winning solution obtained an area under the receiver operating characteristic curve for discerning between severe and non-severe COVID-19 of 0.815. The Final phase solutions of all finalists improved upon their Qualification phase solutions.

Deep learning classification models for medical image analysis often perform well on data from scanners that were used to acquire the training data. However, when these models are applied to data from different vendors, their performance tends to drop substantially. Artifacts that only occur within scans from specific scanners are major causes of this poor generalizability. We aimed to enhance the reliability of deep learning classification models using a novel method called Uncertainty-Based Instance eXclusion (UBIX). UBIX is an inference-time module that can be employed in multiple-instance learning (MIL) settings. MIL is a paradigm in which instances (generally crops or slices) of a bag (generally an image) contribute towards a bag-level output. Instead of assuming equal contribution of all instances to the bag-level output, UBIX detects instances corrupted due to local artifacts on-the-fly using uncertainty estimation, reducing or fully ignoring their contributions before MIL pooling. In our experiments, instances are 2D slices and bags are volumetric images, but alternative definitions are also possible. Although UBIX is generally applicable to diverse classification tasks, we focused on the staging of age-related macular degeneration in optical coherence tomography. Our models were trained on data from a single scanner and tested on external datasets from different vendors, which included vendor-specific artifacts. UBIX showed reliable behavior, with a slight decrease in performance (a decrease of the quadratic weighted kappa (κw) from 0.861 to 0.708), when applied to images from different vendors containing artifacts; while a state-of-the-art 3D neural network without UBIX suffered from a significant detriment of performance (κw from 0.852 to 0.084) on the same test set. We showed that instances with unseen artifacts can be identified with OOD detection. UBIX can reduce their contribution to the bag-level predictions, improving reliability without retraining on new data. This potentially increases the applicability of artificial intelligence models to data from other scanners than the ones for which they were developed. The source code for UBIX, including trained model weights, is publicly available through https://github.com/qurAI-amsterdam/ubix-for-reliable-classification.

We present a novel graph-based approach for labeling the anatomical branches of a given airway tree segmentation. The proposed method formulates airway labeling as a branch classification problem in the airway tree graph, where branch features are extracted using convolutional neural networks and enriched using graph neural networks. Our graph neural network is structure-aware by having each node aggregate information from its local neighbors and position-aware by encoding node positions in the graph. We evaluated the proposed method on 220 airway trees from subjects with various severity stages of Chronic Obstructive Pulmonary Disease (COPD). The results demonstrate that our approach is computationally efficient and significantly improves branch classification performance than the baseline method. The overall average accuracy of our method reaches 91.18% for labeling 18 segmental airway branches, compared to 83.83% obtained by the standard CNN method and 87.37% obtained by the existing method. Furthermore, the reader study done on an additional set of 40 subjects shows that our algorithm performs comparably to human experts in labeling segmental-airways. We published our source code at https://github.com/DIAGNijmegen/spgnn. The proposed algorithm is also publicly available at https://grand-challenge.org/algorithms/airway-anatomical-labeling/.

Pulmonary nodules may be an early manifestation of lung cancer, the leading cause of cancer-related deaths among both men and women. Numerous studies have established that deep learning methods can yield high-performance levels in the detection of lung nodules in chest X-rays. However, the lack of gold-standard public datasets slows down the progression of the research and prevents benchmarking of methods for this task. To address this, we organized a public research challenge, NODE21, aimed at the detection and generation of lung nodules in chest X-rays. While the detection track assesses state-of-the-art nodule detection systems, the generation track determines the utility of nodule generation algorithms to augment training data and hence improve the performance of the detection systems. This paper summarizes the results of the NODE21 challenge and performs extensive additional experiments to examine the impact of the synthetically generated nodule training images on the detection algorithm performance.

The early detection of glaucoma is essential in preventing visual impairment. Artificial intelligence (AI) can be used to analyze color fundus photographs (CFPs) in a cost-effective manner, making glaucoma screening more accessible. While AI models for glaucoma screening from CFPs have shown promising results in laboratory settings, their performance decreases significantly in real-world scenarios due to the presence of out-of-distribution and low-quality images. To address this issue, we propose the Artificial Intelligence for Robust Glaucoma Screening (AIROGS) challenge. This challenge includes a large dataset of around 113,000 images from about 60,000 patients and 500 different screening centers, and encourages the development of algorithms that are robust to ungradable and unexpected input data. We evaluated solutions from 14 teams in this paper and found that the best teams performed similarly to a set of 20 expert ophthalmologists and optometrists. The highest-scoring team achieved an area under the receiver operating characteristic curve of 0.99 (95% CI: 0.98-0.99) for detecting ungradable images on-the-fly. Additionally, many of the algorithms showed robust performance when tested on three other publicly available datasets. These results demonstrate the feasibility of robust AI-enabled glaucoma screening.

Recently, large, high-quality public datasets have led to the development of convolutional neural networks that can detect lymph node metastases of breast cancer at the level of expert pathologists. Many cancers, regardless of the site of origin, can metastasize to lymph nodes. However, collecting and annotating high-volume, high-quality datasets for every cancer type is challenging. In this paper we investigate how to leverage existing high-quality datasets most efficiently in multi-task settings for closely related tasks. Specifically, we will explore different training and domain adaptation strategies, including prevention of catastrophic forgetting, for breast, colon and head-and-neck cancer metastasis detection in lymph nodes. Our results show state-of-the-art performance on colon and head-and-neck cancer metastasis detection tasks. We show the effectiveness of adaptation of networks from one cancer type to another to obtain multi-task metastasis detection networks. Furthermore, we show that leveraging existing high-quality datasets can significantly boost performance on new target tasks and that catastrophic forgetting can be effectively mitigated.Last, we compare different mitigation strategies.

In this work, we report the set-up and results of the Liver Tumor Segmentation Benchmark (LiTS), which was organized in conjunction with the IEEE International Symposium on Biomedical Imaging (ISBI) 2017 and the International Conferences on Medical Image Computing and Computer-Assisted Intervention (MICCAI) 2017 and 2018. The image dataset is diverse and contains primary and secondary tumors with varied sizes and appearances with various lesion-to-background levels (hyper-/hypo-dense), created in collaboration with seven hospitals and research institutions. Seventy-five submitted liver and liver tumor segmentation algorithms were trained on a set of 131 computed tomography (CT) volumes and were tested on 70 unseen test images acquired from different patients. We found that not a single algorithm performed best for both liver and liver tumors in the three events. The best liver segmentation algorithm achieved a Dice score of 0.963, whereas, for tumor segmentation, the best algorithms achieved Dices scores of 0.674 (ISBI 2017), 0.702 (MICCAI 2017), and 0.739 (MICCAI 2018). Retrospectively, we performed additional analysis on liver tumor detection and revealed that not all top-performing segmentation algorithms worked well for tumor detection. The best liver tumor detection method achieved a lesion-wise recall of 0.458 (ISBI 2017), 0.515 (MICCAI 2017), and 0.554 (MICCAI 2018), indicating the need for further research. LiTS remains an active benchmark and resource for research, e.g., contributing the liver-related segmentation tasks in http://medicaldecathlon.com/. In addition, both data and online evaluation are accessible via https://competitions.codalab.org/competitions/17094.

Automatic lesion segmentation on thoracic CT enables rapid quantitative analysis of lung involvement in COVID-19 infections. However, obtaining a large amount of voxel-level annotations for training segmentation networks is prohibitively expensive. Therefore, we propose a weakly-supervised segmentation method based on dense regression activation maps (dRAMs). Most weakly-supervised segmentation approaches exploit class activation maps (CAMs) to localize objects. However, because CAMs were trained for classification, they do not align precisely with the object segmentations. Instead, we produce high-resolution activation maps using dense features from a segmentation network that was trained to estimate a per-lobe lesion percentage. In this way, the network can exploit knowledge regarding the required lesion volume. In addition, we propose an attention neural network module to refine dRAMs, optimized together with the main regression task. We evaluated our algorithm on 90 subjects. Results show our method achieved 70.2% Dice coefficient, substantially outperforming the CAM-based baseline at 48.6%. We published our source code at https://github.com/DIAGNijmegen/bodyct-dram.

An important limitation of state-of-the-art deep learning networks is that they do not recognize when their input is dissimilar to the data on which they were trained and proceed to produce outputs that will be unreliable or nonsensical. In this work, we describe FRODO (Free Rejection of Out-of-Distribution), a publicly available method that can be easily employed for any trained network to detect input data from a different distribution than is expected. FRODO uses the statistical distribution of intermediate layer outputs to define the expected in-distribution (ID) input image properties. New samples are judged based on the Mahalanobis distance (MD) of their layer outputs from the defined distribution. The method can be applied to any network, and we demonstrate the performance of FRODO in correctly rejecting OOD samples on three distinct architectures for classification, localization, and segmentation tasks in chest X-rays. A dataset of 21,576 X-ray images with 3,655 in-distribution samples is defined for testing. The remaining images are divided into four OOD categories of varying levels of difficulty, and performance at rejecting each type is evaluated using receiver operating characteristic (ROC) analysis. FRODO achieves areas under the ROC (AUC) of between 0.815 and 0.999 in distinguishing OOD samples of different types. This is shown to be comparable with the best-performing state-of-the-art method tested, with the substantial advantage that FRODO integrates seamlessly with any network and requires no extra model to be constructed and trained.

Image registration is a fundamental medical image analysis task, and a wide variety of approaches have been proposed. However, only a few studies have comprehensively compared medical image registration approaches on a wide range of clinically relevant tasks. This limits the development of registration methods, the adoption of research advances into practice, and a fair benchmark across competing approaches. The Learn2Reg challenge addresses these limitations by providing a multi-task medical image registration data set for comprehensive characterisation of deformable registration algorithms. A continuous evaluation will be possible at https://learn2reg.grand-challenge.org. Learn2Reg covers a wide range of anatomies (brain, abdomen, and thorax), modalities (ultrasound, CT, MR), availability of annotations, as well as intra- and inter-patient registration evaluation. We established an easily accessible framework for training and validation of 3D registration methods, which enabled the compilation of results of over 65 individual method submissions from more than 20 unique teams. We used a complementary set of metrics, including robustness, accuracy, plausibility, and runtime, enabling unique insight into the current state-of-the-art of medical image registration. This paper describes datasets, tasks, evaluation methods and results of the challenge, as well as results of further analysis of transferability to new datasets, the importance of label supervision, and resulting bias. While no single approach worked best across all tasks, many methodological aspects could be identified that push the performance of medical image registration to new state-of-the-art performance. Furthermore, we demystified the common belief that conventional registration methods have to be much slower than deep-learning-based methods.

Due to memory constraints on current hardware, most convolution neural networks (CNN) are trained on sub-megapixel images. For example, most popular datasets in computer vision contain images much less than a megapixel in size (0.09MP for ImageNet and 0.001MP for CIFAR-10). In some domains such as medical imaging, multi-megapixel images are needed to identify the presence of disease accurately. We propose a novel method to directly train convolutional neural networks using any input image size end-to-end. This method exploits the locality of most operations in modern convolutional neural networks by performing the forward and backward pass on smaller tiles of the image. In this work, we show a proof of concept using images of up to 66-megapixels (8192×8192), saving approximately 50GB of memory per image. Using two public challenge datasets, we demonstrate that CNNs can learn to extract relevant information from these large images and benefit from increasing resolution. We improved the area under the receiver-operating characteristic curve from 0.580 (4MP) to 0.706 (66MP) for metastasis detection in breast cancer (CAMELYON17). We also obtained a Spearman correlation metric approaching state-of-the-art performance on the TUPAC16 dataset, from 0.485 (1MP) to 0.570 (16MP). Code to reproduce a subset of the experiments is available at https://github.com/DIAGNijmegen/StreamingCNN.

Adversarial attacks are considered a potentially serious security threat for machine learning systems. Medical image analysis (MedIA) systems have recently been argued to be vulnerable to adversarial attacks due to strong financial incentives and the associated technological infrastructure. In this paper, we study previously unexplored factors affecting adversarial attack vulnerability of deep learning MedIA systems in three medical domains: ophthalmology, radiology, and pathology. We focus on adversarial black-box settings, in which the attacker does not have full access to the target model and usually uses another model, commonly referred to as surrogate model, to craft adversarial examples that are then transferred to the target model. We consider this to be the most realistic scenario for MedIA systems. Firstly, we study the effect of weight initialization (pre-training on ImageNet or random initialization) on the transferability of adversarial attacks from the surrogate model to the target model, i.e., how effective attacks crafted using the surrogate model are on the target model. Secondly, we study the influence of differences in development (training and validation) data between target and surrogate models. We further study the interaction of weight initialization and data differences with differences in model architecture. All experiments were done with a perturbation degree tuned to ensure maximal transferability at minimal visual perceptibility of the attacks. Our experiments show that pre-training may dramatically increase the transferability of adversarial examples, even when the target and surrogate’s architectures are different: the larger the performance gain using pre-training, the larger the transferability. Differences in the development data between target and surrogate models considerably decrease the performance of the attack; this decrease is further amplified by difference in the model architecture. We believe these factors should be considered when developing security-critical MedIA systems planned to be deployed in clinical practice. We recommend avoiding using only standard components, such as pre-trained architectures and publicly available datasets, as well as disclosure of design specifications, in addition to using adversarial defense methods. When evaluating the vulnerability of MedIA systems to adversarial attacks, various attack scenarios and target-surrogate differences should be simulated to achieve realistic robustness estimates. The code and all trained models used in our experiments are publicly available.3

Recent advances in deep learning have led to a promising performance in many medical image analysis tasks. As the most commonly performed radiological exam, chest radiographs are a particularly important modality for which a variety of applications have been researched. The release of multiple, large, publicly available chest X-ray datasets in recent years has encouraged research interest and boosted the number of publications. In this paper, we review all studies using deep learning on chest radiographs published before March 2021, categorizing works by task: image-level prediction (classification and regression), segmentation, localization, image generation and domain adaptation. Detailed descriptions of all publicly available datasets are included and commercial systems in the field are described. A comprehensive discussion of the current state of the art is provided, including caveats on the use of public datasets, the requirements of clinically useful systems and gaps in the current literature.

Deep-learning-based registration methods emerged as a fast alternative to conventional registration methods. However, these methods often still cannot achieve the same performance as conventional registration methods because they are either limited to small deformation or they fail to handle a superposition of large and small deformations without producing implausible deformation fields with foldings inside. In this paper, we identify important strategies of conventional registration methods for lung registration and successfully developed the deep-learning counterpart. We employ a Gaussian-pyramid-based multilevel framework that can solve the image registration optimization in a coarse-to-fine fashion. Furthermore, we prevent foldings of the deformation field and restrict the determinant of the Jacobian to physiologically meaningful values by combining a volume change penalty with a curvature regularizer in the loss function. Keypoint correspondences are integrated to focus on the alignment of smaller structures. We perform an extensive evaluation to assess the accuracy, the robustness, the plausibility of the estimated deformation fields, and the transferability of our registration approach. We show that it achieves state-of-the-art results on the COPDGene dataset compared to conventional registration method with much shorter execution time. In our experiments on the DIRLab exhale to inhale lung registration, we demonstrate substantial improvements (TRE below 1.2 mm) over other deep learning methods. Our algorithm is publicly available at https://grand-challenge.org/algorithms/deep-learning-based-ct-lung-registration/.

Since its renaissance, deep learning has been widely used in various medical imaging tasks and has achieved remarkable success in many medical imaging applications, thereby propelling us into the so-called artificial intelligence (AI) era. It is known that the success of AI is mostly attributed to the availability of big data with annotations for a single task and the advances in high performance computing. However, medical imaging presents unique challenges that confront deep learning approaches. In this survey paper, we first present traits of medical imaging, highlight both clinical needs and technical challenges in medical imaging, and describe how emerging trends in deep learning are addressing these issues. We cover the topics of network architecture, sparse and noisy labels, federating learning, interpretability, uncertainty quantification, etc. Then, we present several case studies that are commonly found in clinical practice, including digital pathology and chest, brain, cardiovascular, and abdominal imaging. Rather than presenting an exhaustive literature survey, we instead describe some prominent research highlights related to these case study applications. We conclude with a discussion and presentation of promising future directions.

The number of biomedical image analysis challenges organized per year is steadily increasing. These international competitions have the purpose of benchmarking algorithms on common data sets, typically to identify the best method for a given problem. Recent research, however, revealed that common practice related to challenge reporting does not allow for adequate interpretation and reproducibility of results. To address the discrepancy between the impact of challenges and the quality (control), the Biomedical Image Analysis ChallengeS (BIAS) initiative developed a set of recommendations for the reporting of challenges. The BIAS statement aims to improve the transparency of the reporting of a biomedical image analysis challenge regardless of field of application, image modality or task category assessed. This article describes how the BIAS statement was developed and presents a checklist which authors of biomedical image analysis challenges are encouraged to include in their submission when giving a paper on a challenge into review. The purpose of the checklist is to standardize and facilitate the review process and raise interpretability and reproducibility of challenge results by making relevant information explicit.

The triage of acute stroke patients is increasingly dependent on four-dimensional CTA (4D-CTA) imaging. In this work, we present a convolutional neural network (CNN) for image-level detection of intracranial anterior circulation artery occlusions in 4D-CTA. The method uses a normalized 3D time-to-signal (TTS) representation of the input image, which is sensitive to differences in the global arrival times caused by the potential presence of vascular pathologies. The TTS map presents the time within the cranial cavity at which the signal reaches a percentage of the maximum signal intensity, corrected for the baseline intensity. The method was trained and validated on (n=214) patient images and tested on an independent set of (n=279) patient images. This test set included all consecutive suspected-stroke patients admitted to our hospital in 2018. The accuracy, sensitivity, and specificity were 92%, 95%, and 92%. The area under the receiver operating characteristics curve was 0.98 (95% CI: 0.95- 0.99). These results show the feasibility of automated stroke triage in 4D-CTA.

Interpretability of deep learning (DL) systems is gaining attention in medical imaging to increase experts' trust in the obtained predictions and facilitate their integration in clinical settings. We propose a deep visualization method to generate interpretability of DL classification tasks in medical imaging by means of visual evidence augmentation. The proposed method iteratively unveils abnormalities based on the prediction of a classifier trained only with image-level labels. For each image, initial visual evidence of the prediction is extracted with a given visual attribution technique. This provides localization of abnormalities that are then removed through selective inpainting. We iteratively apply this procedure until the system considers the image as normal. This yields augmented visual evidence, including less discriminative lesions which were not detected at first but should be considered for final diagnosis. We apply the method to grading of two retinal diseases in color fundus images: diabetic retinopathy (DR) and age-related macular degeneration (AMD). We evaluate the generated visual evidence and the performance of weakly-supervised localization of different types of DR and AMD abnormalities, both qualitatively and quantitatively. We show that the augmented visual evidence of the predictions highlights the biomarkers considered by experts for diagnosis and improves the final localization performance. It results in a relative increase of 11.2± 2.0% per image regarding sensitivity averaged at 10 false positives/image on average, when applied to different classification tasks, visual attribution techniques and network architectures. This makes the proposed method a useful tool for exhaustive visual support of DL classifiers in medical imaging.

The imaging workup in acute stroke can be simplified by deriving non-contrast CT (NCCT) from CT perfusion (CTP) images. This results in reduced workup time and radiation dose. To achieve this, we present a stacked bidirectional convolutional LSTM (C-LSTM) network to predict 3D volumes from 4D spatiotemporal data. Several parameterizations of the C-LSTM network were trained on a set of 17 CTP-NCCT pairs to learn to derive a NCCT from CTP and were subsequently quantitatively evaluated on a separate cohort of 16 cases. The results show that the C-LSTM network clearly outperforms the baseline and competitive convolutional neural network methods. We show good scalability and performance of the method by continued training and testing on an independent dataset which includes pathology of 80 and 83 CTP-NCCT pairs, respectively. C-LSTM is, therefore, a promising general deep learning approach to learn from high-dimensional spatiotemporal medical images.

Pulmonary lobe segmentation in computed tomography scans is essential for regional assessment of pulmonary diseases. Recent works based on convolution neural networks have achieved good performance for this task. However, they are still limited in capturing structured relationships due to the nature of convolution. The shape of the pulmonary lobes affect each other and their borders relate to the appearance of other structures, such as vessels, airways, and the pleural wall. We argue that such structural relationships play a critical role in the accurate delineation of pulmonary lobes when the lungs are affected by diseases such as COVID-19 or COPD. In this paper, we propose a relational approach (RTSU-Net) that leverages structured relationships by introducing a novel non-local neural network module. The proposed module learns both visual and geometric relationships among all convolution features to produce self-attention weights. With a limited amount of training data available from COVID-19 subjects, we initially train and validate RTSU-Net on a cohort of 5000 subjects from the COPDGene study (4000 for training and 1000 for evaluation). Using models pre-trained on COPDGene, we apply transfer learning to retrain and evaluate RTSU-Net on 470 COVID-19 suspects (370 for retraining and 100 for evaluation). Experimental results show that RTSU-Net outperforms three baselines and performs robustly on cases with severe lung infection due to COVID-19.

Automated detection of cancer metastases in lymph nodes has the potential to improve the assessment of prognosis for patients. To enable fair comparison between the algorithms for this purpose, we set up the CAMELYON17 challenge in conjunction with the IEEE International Symposium on Biomedical Imaging 2017 Conference in Melbourne. Over 300 participants registered on the challenge website, of which 23 teams submitted a total of 37 algorithms before the initial deadline. Participants were provided with 899 whole-slide images (WSIs) for developing their algorithms. The developed algorithms were evaluated based on the test set encompassing 100 patients and 500 WSIs. The evaluation metric used was a quadratic weighted Cohen's kappa. We discuss the algorithmic details of the 10 best pre-conference and two post-conference submissions. All these participants used convolutional neural networks in combination with pre- and postprocessing steps. Algorithms differed mostly in neural network architecture, training strategy, and pre- and postprocessing methodology. Overall, the kappa metric ranged from 0.89 to -0.13 across all submissions. The best results were obtained with pre-trained architectures such as ResNet. Confusion matrix analysis revealed that all participants struggled with reliably identifying isolated tumor cells, the smallest type of metastasis, with detection rates below 40%. Qualitative inspection of the results of the top participants showed categories of false positives, such as nerves or contamination, which could be targeted for further optimization. Last, we show that simple combinations of the top algorithms result in higher kappa metric values than any algorithm individually, with 0.93 for the best combination.

Heavy smokers undergoing screening with low-dose chest CT are affected by cardiovascular disease as much as by lung cancer. Low-dose chest CT scans acquired in screening enable quantification of atherosclerotic calcifications and thus enable identification of subjects at increased cardiovascular risk. This paper presents a method for automatic detection of coronary artery, thoracic aorta, and cardiac valve calcifications in low-dose chest CT using two consecutive convolutional neural networks. The first network identifies and labels potential calcifications according to their anatomical location and the second network identifies true calcifications among the detected candidates. This method was trained and evaluated on a set of 1744 CT scans from the National Lung Screening Trial. To determine whether any reconstruction or only images reconstructed with soft tissue filters can be used for calcification detection, we evaluated the method on soft and medium/sharp filter reconstructions separately. On soft filter reconstructions, the method achieved F1scores of 0.89, 0.89, 0.67, and 0.55 for coronary artery, thoracic aorta, aortic valve, and mitral valve calcifications, respectively. On sharp filter reconstructions, the F1scores were 0.84, 0.81, 0.64, and 0.66, respectively. Linearly weighted kappa coefficients for risk category assignment based on per subject coronary artery calcium were 0.91 and 0.90 for soft and sharp filter reconstructions, respectively. These results demonstrate that the presented method enables reliable automatic cardiovascular risk assessment in all low-dose chest CT scans acquired for lung cancer screening.

Content-based image retrieval (CBIR) is a search technology that could aid medical diagnosis by retrieving and presenting earlier reported cases that are related to the one being diagnosed. To retrieve relevant cases, CBIR systems depend on supervised learning to map low-level image contents to high-level diagnostic concepts. However, the annotation by medical doctors for training and evaluation purposes is a difficult and time-consuming task, which restricts the supervised learning phase to specific CBIR problems of well-defined clinical applications. This paper proposes a new technique that automatically learns the similarity between the several exams from textual distances extracted from radiology reports, thereby successfully reducing the number of annotations needed. Our method first infers the relation between patients by using information retrieval techniques to determine the textual distances between patient radiology reports. These distances are subsequently used to supervise a metric learning algorithm, that transforms the image space accordingly to textual distances. CBIR systems with different image descriptions and different levels of medical annotations were evaluated, with and without supervision from textual distances, using a database of computer tomography scans of patients with interstitial lung diseases. The proposed method consistently improves CBIR mean average precision, with improvements that can reach 38%, and more marked gains for small annotation sets. Given the overall availability of radiology reports in picture archiving and communication systems, the proposed approach can be broadly applied to CBIR systems in different medical problems, and may facilitate the introduction of CBIR in clinical practice.

The papers in this special section focus on the technology and applications supported by deep learning. Deep learning is a growing trend in general data analysis and has been termed one of the 10 breakthrough technologies of 2013. Deep learning is an improvement of artificial neural networks, consisting of more layers that permit higher levels of abstraction and improved predictions from data. To date, it is emerging as the leading machine-learning tool in the general imaging and computer vision domains. In particular, convolutional neural networks (CNNs) have proven to be powerful tools for a broad range of computer vision tasks. Deep CNNs automatically learn mid-level and high-level abstractions obtained from raw data (e.g., images). Recent results indicate that the generic descriptors extracted from CNNs are extremely effective in object recognition and localization in natural images. Medical image analysis groups across the world are quickly entering the field and applying CNNs and other deep learning methodologies to a wide variety of applications.

Convolutional neural networks (CNNs) are deep learning network architectures that have pushed forward the state-of-the-art in a range of computer vision applications and are increasingly popular in medical image analysis. However, training of CNNs is time-consuming and challenging. In medical image analysis tasks, the majority of training examples are easy to classify and therefore contribute little to the CNN learning process. In this paper, we propose a method to improve and speed-up the CNN training for medical image analysis tasks by dynamically selecting misclassified negative samples during training. Training samples are heuristically sampled based on classification by the current status of the CNN. Weights are assigned to the training samples and informative samples are more likely to be included in the next CNN training iteration. We evaluated and compared our proposed method by training a CNN with (SeS) and without (NSeS) the selective sampling method. We focus on the detection of hemorrhages in color fundus images. A decreased training time from 170 epochs to 60 epochs with an increased performance-on par with two human experts-was achieved with areas under the receiver operating characteristics curve of 0.894 and 0.972 on two data sets. The SeS CNN statistically outperformed the NSeS CNN on an independent test set.

The major advantage of multiple-instance learning (MIL) applied to a computer-aided detection (CAD) system is that it allows optimizing the latter with case-level labels instead of accurate lesion outlines as traditionally required for a supervised approach. As shown in previous work, a MIL-based CAD system can perform comparably to its supervised counterpart considering complex tasks such as chest radiograph scoring in tuberculosis (TB) detection. However, despite this remarkable achievement, the uncertainty inherent to MIL can lead to a less satisfactory outcome if analysis at lower levels (e.g., regions or pixels) is needed. This issue may seriously compromise the applicability of MIL to tasks related to quantification or grading, or detection of highly localized lesions. In this paper, we propose to reduce uncertainty by embedding a MIL classifier within an active learning (AL) framework. To minimize the labeling effort, we develop a novel instance selection mechanism that exploits the MIL problem definition through one-class classification. We adapt this mechanism to provide meaningful regions instead of individual instances for expert labeling, which is a more appropriate strategy given the application domain. In addition, and contrary to usual AL methods, a single iteration is performed. To show the effectiveness of our approach, we compare the output of a MIL-based CAD system trained with and without the proposed AL framework. The task is to detect textural abnormalities related to TB. Both quantitative and qualitative evaluations at the pixel level are carried out. Our method significantly improves the MIL-based classification.

We propose a novel Computer-Aided Detection (CAD) system for pulmonary nodules using multi-view convolutional networks (ConvNets), for which discriminative features are automatically learnt from the training data. The network is fed with nodule candidates obtained by combining three candidate detectors specifically designed for solid, subsolid, and large nodules. For each candidate, a set of 2-D patches from differently oriented planes is extracted. The proposed architecture comprises multiple streams of 2-D ConvNets, for which the outputs are combined using a dedicated fusion method to get the final classification. Data augmentation and dropout are applied to avoid overfitting. On 888 scans of the publicly available LIDC-IDRI dataset, our method reaches high detection sensitivities of 85.4% and 90.1% at 1 and 4 false positives per scan, respectively. An additional evaluation on independent datasets from the ANODE09 challenge and DLCST is performed. We showed that the proposed multi-view ConvNets is highly suited to be used for false positive reduction of a CAD system.

We present a novel descriptor for the characterization of pulmonary nodules in computed tomography (CT) images. The descriptor encodes information on nodule morphology and has scale-invariant and rotation-invariant properties. Information on nodule morphology is captured by sampling intensity profiles along circular patterns on spherical surfaces centered on the nodule, in a multi-scale fashion. Each intensity profile is interpreted as a periodic signal, where the Fourier transform is applied, obtaining a spectrum. A library of spectra is created and labeled via unsupervised clustering, obtaining a Bag-of-Frequencies, which is used to assign each spectra a label. The descriptor is obtained as the histogram of labels along all the spheres. Additional contributions are a technique to estimate the nodule size, based on the sampling strategy, as well as a technique to choose the most informative plane to cut a 2-D view of the nodule in the 3-D image. We evaluate the descriptor on several nodule morphology classification problems, namely discrimination of nodules versus vascular structures and characterization of spiculation. We validate the descriptor on data from European screening trials NELSON and DLCST and we compare it with state-of-the-art approaches for 3-D shape description in medical imaging and computer vision, namely SPHARM and 3-D SIFT, outperforming them in all the considered experiments.

Tuberculosis (TB) is a common disease with high mortality and morbidity rates worldwide. Automatic systems to detect TB on chest radiographs (CXRs) can improve the efficiency of diagnostic algorithms for pulmonary TB. The diverse manifestation of TB on CXRs from different populations requires a system that can be adapted to deal with different types of abnormalities. A computer aided detection (CAD) system was developed which combines several subscores of supervised subsystems detecting textural, shape, and focal abnormalities into one TB score. A general framework was developed to combine an arbitrary number of subscores: subscores were normalized, collected in a feature vector and then combined using a supervised classifier into one combined score. The method was evaluated on two databases, both consisting of 200 digital CXRs, from: (A) Western high-risk group screening, (B) TB suspect screening in Africa. The subscores and combined score were compared to (1) an external, non-radiological, reference and (2) a radiological reference determined by a human expert. Performance was measured using Receiver Operator Characteristic (ROC) analysis. Different subscores performed best in the two databases. The combined TB score performed better than the individual subscores, except for the external reference in database B. The performances of the independent observer were slightly higher than the combined TB score. Compared to the external reference, differences in performance between the combined TB score and the independent observer were not significant in both databases. Supervised combination to compute an overall TB score allows for a necessary adaptation of the CAD system to different settings or different operational requirements.

To reach performance levels comparable to human experts, computer-aided detection (CAD) systems are typically optimized following a supervised learning approach that relies on large training databases comprising manually annotated lesions. However, manually outlining those lesions constitutes a difficult and time-consuming process that renders detailedly annotated data difficult to obtain. In this paper, we investigate an alternative approach, namely multiple-instance learning (MIL), that does not require detailed information for optimization. We have applied MIL to a CAD system for tuberculosis detection. Only the case condition (normal or abnormal) was required during training. Based upon the well-known miSVM technique, we propose an improved algorithm that overcomes miSVM's drawbacks related to positive instance underestimation and costly iteration. To show the advantages of our MIL-based approach as compared with a traditional supervised one, experiments with three X-ray databases were conducted. The area under the receiver operating characteristic curve was utilized as a performance measure. With the first database, for which training lesion annotations were available, our MIL-based method was comparable to the supervised system ( 0.86 versus 0.88 ). When evaluating the remaining databases, given their large difference with the previous image set, the most appealing strategy was to retrain the CAD systems. However, since only the case condition was available, only the MIL-based system could be retrained. This scenario, which is common in real-world applications, demonstrates the better adaptation capabilities of the proposed approach. After retraining, our MIL-based system significantly outperformed the supervised one ( 0.86 versus 0.79 and 0.91 versus 0.85 , and p=0.0002 , respectively).

Automated quantitative analysis systems for medical images often lack the capability to successfully process images from multiple sources. Normalization of such images prior to further analysis is a possible solution to this limitation. This work presents a general method to normalize medical images and thoroughly investigates its effectiveness for chest radiography (CXR). The method starts with an energy decomposition of the image in different bands. Next, each band's localized energy is scaled to a reference value and the image is reconstructed. We investigate iterative and local application of this technique. The normalization is applied iteratively to the lung fields on six datasets from different sources, each comprising 50 normal CXRs and 50 abnormal CXRs. The method is evaluated in three supervised computer-aided detection tasks related to CXR analysis and compared to two reference normalization methods. In the first task, automatic lung segmentation, the average Jaccard overlap significantly increased from 0.72±0.30 and 0.87±0.11 for both reference methods to with normalization. The second experiment was aimed at segmentation of the clavicles. The reference methods had an average Jaccard index of 0.57±0.26 and 0.53±0.26; with normalization this significantly increased to . The third experiment was detection of tuberculosis related abnormalities in the lung fields. The average area under the Receiver Operating Curve increased significantly from 0.72±0.14 and 0.79±0.06 using the reference methods to with normalization. We conclude that the normalization can be successfully applied in chest radiography and makes supervised systems more generally applicable to data from different sources.

Projection images, such as those routinely acquired in radiological practice, are difficult to analyze because multiple 3-D structures superimpose at a single point in the 2-D image. Removal of particular superimposed structures may improve interpretation of these images, both by humans and by computers. This work therefore presents a general method to isolate and suppress structures in 2-D projection images. The focus is on elongated structures, which allows an intensity model of a structure of interest to be extracted using local information only. The model is created from profiles sampled perpendicular to the structure. Profiles containing other structures are detected and removed to reduce the influence on the model. Subspace filtering, using blind source separation techniques, is applied to separate the structure to be suppressed from other structures. By subtracting the modeled structure from the original image a structure suppressed image is created. The method is evaluated in four experiments. In the first experiment ribs are suppressed in 20 artificial radiographs simulated from 3-D lung computed tomography (CT) images. The proposed method with blind source separation and outlier detection shows superior suppression of ribs in simulated radiographs, compared to a simplified approach without these techniques. Additionally, the ability of three observers to discriminate between patches containing ribs and containing no ribs, as measured by the area under the receiver operating characteristic curve (AUC), reduced from 0.99-1.00 on original images to 0.75-0.84 on suppressed images. In the second experiment clavicles are suppressed in 253 chest radiographs. The effect of suppression on clavicle visibility is evaluated using the clavicle contrast and border response, showing a reduction of 78% and 34%, respectively. In the third experiment nodules extracted from CT were simulated close to the clavicles in 100 chest radiographs. It was found that after suppression contrast of the nodules was higher than of the clavicles (1.35 and 0.55, respectively) than on original images (1.83 and 2.46, respectively). In the fourth experiment catheters were suppressed in chest radiographs. The ability of three observers to discriminate between patches originating from 36 images with and 21 images without catheters, as measured by the AUC, reduced from 0.98-0.99 on original images to 0.64-0.74 on suppressed images. We conclude that the presented method can markedly reduce the visibility of elongated structures in chest radiographs and shows potential to enhance diagnosis.

Segmentation of the pulmonary lobes is relevant in clinical practice and particularly challenging for cases with severe diseases or incomplete fissures. In this work, an automated segmentation approach is presented that performs a marker-based watershed transformation on computed tomography (CT) scans to subdivide the lungs into lobes. A cost image for the watershed transformation is computed by combining information from fissures, bronchi, and pulmonary vessels. The lobar markers are calculated by an analysis of the automatically labeled bronchial tree. By integration of information from several anatomical structures the segmentation is made robust against incomplete fissures. For evaluation the method was compared to a recently published method on 20 CT scans with no or mild disease. The average distances to the reference segmentation were 0.69, 0.67, and 1.21 mm for the left major, right major, and right minor fissure, respectively. In addition the results were submitted to LOLA11, an international lung lobe segmentation challenge with publically available data including cases with severe diseases. The average distances to the reference for the 55 CT scans provided by LOLA11 were 0.98, 3.97, and 3.09 mm for the left major, right major, and right minor fissure. Moreover, an analysis of the relation between segmentation quality and fissure completeness showed that the method is robust against incomplete fissures.

The calcium burden as estimated from non-ECG-synchronized computed tomography (CT) exams acquired in screening of heavy smokers has been shown to be a strong predictor of cardiovascular events. We present a method for automatic coronary calcium scoring with low-dose, non-contrast-enhanced, non-ECG-synchronized chest CT. First, a probabilistic coronary calcium map was created using multi-atlas segmentation. This map assigned an a priori probability for the presence of coronary calcifications at every location in a scan. Subsequently, a statistical pattern recognition system was designed to identify coronary calcifications by texture, size, and spatial features; the spatial features were computed using the coronary calcium map. The detected calcifications were quantified in terms of volume and Agatston score. The best results were obtained by merging the results of three different supervised classification systems, namely direct classification with a nearest neighbor classifier, and two-stage classification with nearest neighbor and support vector machine classifiers.We used a total of 231 test scans containing 45,674 mm³ of coronary calcifications. The presented method detected on average 157/198 mm³ (sensitivity 79.2%) of coronary calcium volume with on average 4 mm false positive volume. Calcium scoring can be performed automatically in low-dose, non-contrast enhanced, non-ECG-synchronized chest CT in screening of heavy smokers to identify subjects who might benefit from preventive treatment.

This paper describes a framework for establishing a reference airway tree segmentation, which was used to quantitatively evaluate fifteen different airway tree extraction algorithms in a standardized manner. Because of the sheer difficulty involved in manually constructing a complete reference standard from scratch, we propose to construct the reference using results from all algorithms that are to be evaluated. We start by subdividing each segmented airway tree into its individual branch segments. Each branch segment is then visually scored by trained observers to determine whether or not it is a correctly segmented part of the airway tree. Finally, the reference airway trees are constructed by taking the union of all correctly extracted branch segments. Fifteen airway tree extraction algorithms from different research groups are evaluated on a diverse set of twenty chest computed tomography (CT) scans of subjects ranging from healthy volunteers to patients with severe pathologies, scanned at different sites, with different CT scanner brands, models, and scanning protocols. Three performance measures covering different aspects of segmentation quality were computed for all participating algorithms. Results from the evaluation showed that no single algorithm could extract more than an average of 74% of the total length of all branches in the reference standard, indicating substantial differences between the algorithms. A fusion scheme that obtained superior results is presented, demonstrating that there is complementary information provided by the different algorithms and there is still room for further improvements in airway segmentation algorithms.

EMPIRE10 (Evaluation of Methods for Pulmonary Image REgistration 2010) is a public platform for fair and meaningful comparison of registration algorithms which are applied to a database of intrapatient thoracic CT image pairs. Evaluation of nonrigid registration techniques is a nontrivial task. This is compounded by the fact that researchers typically test only on their own data, which varies widely. For this reason, reliable assessment and comparison of different registration algorithms has been virtually impossible in the past. In this work we present the results of the launch phase of EMPIRE10, which comprised the comprehensive evaluation and comparison of 20 individual algorithms from leading academic and industrial research groups. All algorithms are applied to the same set of 30 thoracic CT pairs. Algorithm settings and parameters are chosen by researchers expert in the configuration of their own method and the evaluation is independent, using the same criteria for all participants. All results are published on the EMPIRE10 website (http://empire10.isi.uu.nl). The challenge remains ongoing and open to new participants. Full results from 24 algorithms have been published at the time of writing. This paper details the organization of the challenge, the data and evaluation methods and the outcome of the initial launch with 20 algorithms. The gain in knowledge and future work are discussed.

Computer-aided detection (CAD) is increasingly used in clinical practice and for many applications a multitude of CAD systems have been developed. In practice, CAD systems have different strengths and weaknesses and it is therefore interesting to consider their combination. In this paper, we present generic methods to combine multiple CAD systems and investigate what kind of performance increase can be expected. Experimental results are presented using data from the ANODE09 and ROC09 online CAD challenges for the detection of pulmonary nodules in computed tomography scans and red lesions in retinal images, respectively. For both applications, combination results in a large and significant increase in performance when compared to the best individual CAD system.

A decreased ratio of the width of retinal arteries to veins [arteriolar-to-venular diameter ratio (AVR)], is well established as predictive of cerebral atrophy, stroke and other cardiovascular events in adults. Tortuous and dilated arteries and veins, as well as decreased AVR are also markers for plus disease in retinopathy of prematurity. This work presents an automated method to estimate the AVR in retinal color images by detecting the location of the optic disc, determining an appropriate region of interest (ROI), classifying vessels as arteries or veins, estimating vessel widths, and calculating the AVR. After vessel segmentation and vessel width determination, the optic disc is located and the system eliminates all vessels outside the AVR measurement ROI. A skeletonization operation is applied to the remaining vessels after which vessel crossings and bifurcation points are removed, leaving a set of vessel segments consisting of only vessel centerline pixels. Features are extracted from each centerline pixel in order to assign these a soft label indicating the likelihood that the pixel is part of a vein. As all centerline pixels in a connected vessel segment should be the same type, the median soft label is assigned to each centerline pixel in the segment. Next, artery vein pairs are matched using an iterative algorithm, and the widths of the vessels are used to calculate the AVR. We trained and tested the algorithm on a set of 65 high resolution digital color fundus photographs using a reference standard that indicates for each major vessel in the image whether it is an artery or vein. We compared the AVR values produced by our system with those determined by a semi-automated reference system. We obtained a mean unsigned error of 0.06 (SD 0.04) in 40 images with a mean AVR of 0.67. A second observer using the semi-automated system obtained the same mean unsigned error of 0.06 (SD 0.05) on the set of images with a mean AVR of 0.66. The testing data and reference standard used in this study has been made publicly available.

The detection of microaneurysms in digital color fundus photographs is a critical first step in automated screening for diabetic retinopathy (DR), a common complication of diabetes. To accomplish this detection numerous methods have been published in the past but none of these was compared with each other on the same data. In this work we present the results of the first international microaneurysm detection competition, organized in the context of the Retinopathy Online Challenge (ROC), a multiyear online competition for various aspects of DR detection. For this competition, we compare the results of five different methods, produced by five different teams of researchers on the same set of data. The evaluation was performed in a uniform manner using an algorithm presented in this work. The set of data used for the competition consisted of 50 training images with available reference standard and 50 test images where the reference standard was withheld by the organizers (M. Niemeijer, B. van Ginneken, and M. D. Abràmoff). The results obtained on the test data was submitted through a website after which standardized evaluation software was used to determine the performance of each of the methods. A human expert detected microaneurysms in the test set to allow comparison with the performance of the automatic methods. The overall results show that microaneurysm detection is a challenging task for both the automatic methods as well as the human expert. There is room for improvement as the best performing system does not reach the performance of the human expert. The data associated with the ROC microaneurysm detection competition will remain publicly available and the website will continue accepting submissions.

A method for automatic segmentation of pulmonary lobes from computed tomography (CT) scans is presented that is robust against incomplete fissures. The method is based on a multiatlas approach in which existing lobar segmentations are deformed to test scans in which the fissures, the lungs, and the bronchial tree have been automatically segmented. The key element of our method is a cost function that exploits information from fissures, lung borders, and bronchial tree in an effective way, such that less reliable information (lungs, airways) is only used when the most reliable information (fissures) is missing. To cope with the anatomical variation in lobe shape, an atlas selection mechanism is introduced. The method is evaluated on two test sets of 120 scans in total. The results show that the lobe segmentation closely follows the fissures when they are present. In a simulated experiment in which parts of complete fissures are removed, the robustness of the method against different levels of incomplete fissures is shown. When the fissures are incomplete, an observer study shows agreement of the automatically determined lobe borders with a radiologist for 81% of the lobe borders on average.

This paper presents a comparison study between 10 automatic and six interactive methods for liver segmentation from contrast-enhanced CT images. It is based on results from the "MICCAI 2007 Grand Challenge" workshop, where 16 teams evaluated their algorithms on a common database. A collection of 20 clinical images with reference segmentations was provided to train and tune algorithms in advance. Participants were also allowed to use additional proprietary training data for that purpose. All teams then had to apply their methods to 10 test datasets and submit the obtained results. Employed algorithms include statistical shape models, atlas registration, level-sets, graph-cuts and rule-based systems. All results were compared to reference segmentations five error measures that highlight different aspects of segmentation accuracy. All measures were combined according to a specific scoring system relating the obtained values to human expert variability. In general, interactive methods reached higher average scores than automatic approaches and featured a better consistency of segmentation quality. However, the best automatic methods (mainly based on statistical shape models with some additional free deformation) could compete well on the majority of test images. The study provides an insight in performance of different segmentation approaches under real-world conditions and highlights achievements and limitations of current image analysis techniques.

A novel atlas-based segmentation approach based on the combination of multiple registrations is presented. Multiple atlases are registered to a target image. To obtain a segmentation of the target, labels of the atlas images are propagated to it. The propagated labels are combined by spatially varying decision fusion weights. These weights are derived from local assessment of the registration success. Furthermore, an atlas selection procedure is proposed that is equivalent to sequential forward selection from statistical pattern recognition theory. The proposed method is compared to three existing atlas-based segmentation approaches, namely 1) single atlas-based segmentation, 2) average-shape atlas-based segmentation, and 3) multi-atlas-based segmentation with averaging as decision fusion. These methods were tested on the segmentation of the heart and the aorta in computed tomography scans of the thorax. The results show that the proposed method outperforms other methods and yields results very close to those of an independent human observer. Moreover, the additional atlas selection step led to a faster segmentation at a comparable performance.

Noise filtering techniques that maintain image contrast while decreasing image noise have the potential to optimize the quality of computed tomography (CT) images acquired at reduced radiation dose. In this paper, a hybrid diffusion filter with continuous switch (HDCS) is introduced, which exploits the benefits of three-dimensional edge-enhancing diffusion (EED) and coherence-enhancing diffusion (CED). Noise is filtered, while edges, tubular structures, and small spherical structures are preserved. From ten high dose thorax CT scans, acquired at clinical doses, ultra low dose ( 15 mAs ) scans were simulated and used to evaluate and compare HDCS to other diffusion filters, such as regularized Perona-Malik diffusion and EED. Quantitative results show that the HDCS filter outperforms the other filters in restoring the high dose CT scan from the corresponding simulated low dose scan. A qualitative evaluation was performed on filtered real low dose CT thorax scans. An expert observer scored artifacts as well as fine structures and was asked to choose one of three scans (two filtered (blinded), one unfiltered) for three different settings (trachea, lung, and mediastinal). Overall, the HDCS filtered scan was chosen most often.

The purpose of computer-aided detection or diagnosis (CAD) technology has so far been to serve as a second reader. If, however, all relevant lesions in an image can be detected by CAD algorithms, use of CAD for automatic reading or prescreening may become feasible. This work addresses the question how to fuse information from multiple CAD algorithms, operating on multiple images that comprise an exam, to determine a likelihood that the exam is normal and would not require further inspection by human operators. We focus on retinal image screening for diabetic retinopathy, a common complication of diabetes. Current CAD systems are not designed to automatically evaluate complete exams consisting of multiple images for which several detection algorithm output sets are available. Information fusion will potentially play a crucial role in enabling the application of CAD technology to the automatic screening problem. Several different fusion methods are proposed and their effect on the performance of a complete comprehensive automatic diabetic retinopathy screening system is evaluated. Experiments show that the choice of fusion method can have a large impact on system performance. The complete system was evaluated on a set of 15,000 exams (60,000 images). The best performing fusion method obtained an area under the receiver operator characteristic curve of 0.881. This indicates that automated prescreening could be applied in diabetic retinopathy screening programs.

Automated extraction of pulmonary anatomy provides a foundation for computerized analysis of computed tomography (CT) scans of the chest. A completely automatic method is presented to segment the lungs, lobes and pulmonary segments from volumetric CT chest scans. The method starts with lung segmentation based on region growing and standard image processing techniques. Next, the pulmonary fissures are extracted by a supervised filter. Subsequently the lung lobes are obtained by voxel classification where the position of voxels in the lung and relative to the fissures are used as features. Finally, each lobe is subdivided in its pulmonary segments by applying another voxel classification that employs features based on the detected fissures and the relative position of voxels in the lobe. The method was evaluated on 100 low-dose CT scans obtained from a lung cancer screening trial and compared to estimates of both interobserver and intraobserver agreement. The method was able to segment the pulmonary segments with high accuracy (77%), comparable to both interobserver and intraobserver accuracy (74% and 80%, respectively).

In medical image processing, many filters have been developed to enhance certain structures in 3-D data. In this paper, we propose to use pattern recognition techniques to design more optimal filters. The essential difference with previous approaches is that we provide a system with examples of what it should enhance and suppress. This training data is used to construct a classifier that determines the probability that a voxel in an unseen image belongs to the target structure(s). The output of a rich set of basis filters serves as input to the classifier. In a feature selection process, this set is reduced to a compact, efficient subset. We show that the output of the system can be reused to extract new features, using the same filters, that can be processed by a new classifier. Such a multistage approach further improves performance. While the approach is generally applicable, in this work the focus is on enhancing pulmonary fissures in 3-D computed tomography (CT) chest scans. A supervised fissure enhancement filter is evaluated on two data sets, one of scans with a normal clinical dose and one of ultra-low dose scans. Results are compared with those of a recently proposed conventional fissure enhancement filter. It is demonstrated that both methods are able to enhance fissures, but the supervised approach shows better performance; the areas under the receiver operating characteristic (ROC) curve are 0.98 versus 0.90, for the normal dose data and 0.97 versus 0.87 for the ultra low dose data, respectively.

An automatic system is presented to find the location of the major anatomical structures in color fundus photographs; the optic disc, the macula, and the vascular arch. These structures are found by fitting a single point-distribution-model to the image, that contains points on each structure. The method can handle optic disc and macula centered images of both the left and the right eye. The system uses a cost function, which is based on a combination of both global and local cues, to find the correct position of the model points. The global terms in the cost function are based on the orientation and width of the vascular pattern in the image. The local term is derived from the image structure around the points of the model. To optimize the fit of the point-distribution-model to an image, a sophisticated combination of optimization processes is proposed which combines optimization in the parameter space of the model and in the image space, where points are moved directly. Experimental results are presented demonstrating that our specific choices for the cost function components and optimization scheme are needed to obtain good results. The system was developed and trained on a set of 500 screening images, and tested on a completely independent set of 500 screening images. In addition to this the system was also tested on a separate set of 100 pathological images. In the screening set it was able to find the vascular arch in 93.2%, the macula in 94.4%, the optic disc location in 98.4% and whether it is dealing with a left or right eye in 100% of all tested cases. For the pathological images test set, this was 77.0%, 92.0%, 94.0%, and 100% respectively.

The task of segmenting the posterior ribs within the lung fields of standard posteroanterior chest radiographs is considered. To this end, an iterative, pixel-based, supervised, statistical classification method is used, which is called iterated contextual pixel classification (ICPC). Starting from an initial rib segmentation obtained from pixel classification, ICPC updates it by reclassifying every pixel, based on the original features and, additionally, class label information of pixels in the neighborhood of the pixel to be reclassified. The method is evaluated on 30 radiographs taken from the JSRT (Japanese Society of Radiological Technology) database. All posterior ribs within the lung fields in these images have been traced manually by two observers. The first observer's segmentations are set as the gold standard; ICPC is trained using these segmentations. In a sixfold cross-validation experiment, ICPC achieves a classification accuracy of 0.86 +/- 0.06, as compared to 0.94 +/- 0.02 for the second human observer.

Computed tomography (CT) has become the new reference standard for quantification of emphysema. The most popular measure of emphysema derived from CT is the pixel index (PI), which expresses the fraction of the lung volume with abnormally low intensity values. As PI is calculated from a single, fixed threshold on intensity, this measure is strongly influenced by noise. This effect shows up clearly when comparing the PI score of a high-dose scan to the PI score of a low-dose (i.e., noisy) scan of the same subject. In this paper, the noise variance (NOVA) filter is presented: a general framework for (iterative) nonlinear filtering, which uses an estimate of the spatially dependent noise variance in an image. The NOVA filter iteratively estimates the local image noise and filters the image. For the specific purpose of emphysema quantification of low-dose CT images, a dedicated, noniterative NOVA filter is constructed by using prior knowledge of the data to obtain a good estimate of the spatially dependent noise in an image. The performance of the NOVA filter is assessed by comparing characteristics of pairs of high-dose and low-dose scans. The compared characteristics are the PI scores for different thresholds and the size distributions of emphysema bullae. After filtering, the PI scores of high-dose and low-dose images agree to within 2%-3% points. The reproducibility of the high-dose bullae size distribution is also strongly improved. NOVA filtering of a CT image of typically 400 x 512 x 512 voxels takes only a couple of minutes which makes it suitable for routine use in clinical practice.

Current computed tomography (CT) technology allows for near isotropic, submillimeter resolution acquisition of the complete chest in a single breath hold. These thin-slice chest scans have become indispensable in thoracic radiology, but have also substantially increased the data load for radiologists. Automating the analysis of such data is, therefore, a necessity and this has created a rapidly developing research area in medical imaging. This paper presents a review of the literature on computer analysis of the lungs in CT scans and addresses segmentation of various pulmonary structures, registration of chest scans, and applications aimed at detection, classification and quantification of chest abnormalities. In addition, research trends and challenges are identified and directions for future research are discussed.

We present a novel method that determines whether a macula centered retinal image is from the left or right eye and automatically detects the optic disc, the fovea and the vascular arch by inferring the location of a set of landmarks placed on these structures. The algorithm relies on a specific energy function that combines global and local cues. The global cues are derived from vascular atlases of the vessel orientation and thickness on the retina as well as a vascular distance map. A fourth component models the local appearance around each of the landmarks in the model and is able to estimate the distance between a position in the image and the target position of a landmark. For the minimization of the energy function a combination of optimization methods is used. We compare the results of several different system setups and combinations of energy function components with the performance of a second human observer. The best performing system localizes the OD in 91% of all cases, the fovea in 94% of all cases and correctly positions 74% of all vessel landmarks. The results show that a combination of global and local energy function components is required to obtain optimal results.

The robust detection of red lesions in digital color fundus photographs is a critical step in the development of automated screening systems for diabetic retinopathy. In this paper, a novel red lesion detection method is presented based on a hybrid approach, combining prior works by Spencer et al. (1996) and Frame et al. (1998) with two important new contributions. The first contribution is a new red lesion candidate detection system based on pixel classification. Using this technique, vasculature and red lesions are separated from the background of the image. After removal of the connected vasculature the remaining objects are considered possible red lesions. Second, an extensive number of new features are added to those proposed by Spencer-Frame. The detected candidate objects are classified using all features and a k-nearest neighbor classifier. An extensive evaluation was performed on a test set composed of images representative of those normally found in a screening set. When determining whether an image contains red lesions the system achieves a sensitivity of 100% at a specificity of 87%. The method is compared with several different automatic systems and is shown to outperform them all. Performance is close to that of a human expert examining the images for the presence of red lesions.

Conventional methods of lung segmentation rely on a large gray value contrast between lung fields and surrounding tissues. These methods fail on scans with lungs that contain dense pathologies, and such scans occur frequently in clinical practice. We propose a segmentation-by-registration scheme in which a scan with normal lungs is elastically registered to a scan containing pathology. When the resulting transformation is applied to a mask of the normal lungs, a segmentation is found for the pathological lungs. As a mask of the normal lungs, a probabilistic segmentation built up out of the segmentations of 15 registered normal scans is used. To refine the segmentation, voxel classification is applied to a certain volume around the borders of the transformed probabilistic mask. Performance of this scheme is compared to that of three other algorithms: a conventional, a user-interactive and a voxel classification method. The algorithms are tested on 10 three-dimensional thin-slice computed tomography volumes containing high-density pathology. The resulting segmentations are evaluated by comparing them to manual segmentations in terms of volumetric overlap and border positioning measures. The conventional and user-interactive methods that start off with thresholding techniques fail to segment the pathologies and are outperformed by both voxel classification and the refined segmentation-by-registration. The refined registration scheme enjoys the additional benefit that it does not require pathological (hand-segmented) training data.

A method is presented for automated segmentation of vessels in two-dimensional color images of the retina. This method can be used in computer analyses of retinal images, e.g., in automated screening for diabetic retinopathy. The system is based on extraction of image ridges, which coincide approximately with vessel centerlines. The ridges are used to compose primitives in the form of line elements. With the line elements an image is partitioned into patches by assigning each image pixel to the closest line element. Every line element constitutes a local coordinate frame for its corresponding patch. For every pixel, feature vectors are computed that make use of properties of the patches and the line elements. The feature vectors are classified using a kappaNN-classifier and sequential forward feature selection. The algorithm was tested on a database consisting of 40 manually labeled images. The method achieves an area under the receiver operating characteristic curve of 0.952. The method is compared with two recently published rule-based methods of Hoover et al. and Jiang et al. The results show that our method is significantly better than the two rule-based methods (p < 0.01). The accuracy of our method is 0.944 versus 0.947 for a second observer.

An active shape model segmentation scheme is presented that is steered by optimal local features, contrary to normalized first order derivative profiles, as in the original formulation [Cootes and Taylor, 1995, 1999, and 2001]. A nonlinear kNN-classifier is used, instead of the linear Mahalanobis distance, to find optimal displacements for landmarks. For each of the landmarks that describe the shape, at each resolution level taken into account during the segmentation optimization procedure, a distinct set of optimal features is determined. The selection of features is automatic, using the training images and sequential feature forward and backward selection. The new approach is tested on synthetic data and in four medical segmentation tasks: segmenting the right and left lung fields in a database of 230 chest radiographs, and segmenting the cerebellum and corpus callosum in a database of 90 slices from MRI brain images. In all cases, the new method produces significantly better results in terms of an overlap error measure (p < 0.001 using a paired T-test) than the original active shape model scheme.

A fully automatic method is presented to detect abnormalities in frontal chest radiographs which are aggregated into an overall abnormality score. The method is aimed at finding abnormal signs of a diffuse textural nature, such as they are encountered in mass chest screening against tuberculosis (TB). The scheme starts with automatic segmentation of the lung fields, using active shape models. The segmentation is used to subdivide the lung fields into overlapping regions of various sizes. Texture features are extracted from each region, using the moments of responses to a multiscale filter bank. Additional "difference features" are obtained by subtracting feature vectors from corresponding regions in the left and right lung fields. A separate training set is constructed for each region. All regions are classified by voting among the k nearest neighbors, with leave-one-out. Next, the classification results of each region are combined, using a weighted multiplier in which regions with higher classification reliability weigh more heavily. This produces an abnormality score for each image. The method is evaluated on two databases. The first database was collected from a TB mass chest screening program, from which 147 images with textural abnormalities and 241 normal images were selected. Although this database contains many subtle abnormalities, the classification has a sensitivity of 0.86 at a specificity of 0.50 and an area under the receiver operating characteristic (ROC) curve of 0.820. The second database consist of 100 normal images and 100 abnormal images with interstitial disease. For this database, the results were a sensitivity of 0.97 at a specificity of 0.90 and an area under the ROC curve of 0.986.

The traditional chest radiograph is still ubiquitous in clinical practice, and will likely remain so for quite some time. Yet, its interpretation is notoriously difficult. This explains the continued interest in computer-aided diagnosis for chest radiography. The purpose of this survey is to categorize and briefly review the literature on computer analysis of chest images, which comprises over 150 papers published in the last 30 years. Remaining challenges are indicated and some directions for future research are given.

In this work, we investigate the visual appearance of real-world surfaces and the dependence of appearance on the geometry of imaging conditions. We discuss a new texture representation called the BTF (bidirectional texture function) which captures the variation in texture with illumination and viewing direction. We present a BTF database with image textures from over 60 different samples, each observed with over 200 different combinations of viewing and illumination directions. We describe the methods involved in collecting the database as well as the importqance and uniqueness of this database for computer graphics. A related quantity to the BTF is the familiar BRDF (bidirectional reflectance distribution function). The measurement methods involved in the BTF database are conducive to simultaneous measurement of the BRDF. Accordingly, we also present a BRDF database with reflectance measurements for over 60 different samples, each observed with over 200 different combinations of viewing and illumination directions. Both of these unique databases are publicly available and have important implications for computer graphics.

In this work, we investigate the visual appearance of real-world surfaces and the dependence of appearance on imaging conditions. We present a BRDF (bidirectional reflectance distribution function) database with reflectance measurements for over 60 different samples, each observed with over 200 different combinations of viewing and source directions. We fit the BRDF measurements to two recent models to obtain a BRDF parameter database. These BRDF parameters can be directly used for both image analysis and image synthesis. Finally, we present a BTF (bidirectional texture function) database with image textures from over 60 different samples, each observed with over 200 different combinations of viewing and source directions. Each of these unique databases has important implications for a variety of vision algorithms and each is made publicly available.

REALISE was designed to extract from sequences of images, acquired with a moving camera, the information necessary for determining the 3D (CAD-like) structure of a real-life scene together with information about the radiometric signatures of surfaces bounding the extracted 3D objects (e.g. reflectance behaviour). The retrieved information is then integrated in a virtual reality (VR) software environment. The R&D work is been performed principally in the following areas of computer vision and computer graphics: structure from motion, recovery of geometries, recovery of photometric and texture information, highly realistic rendering on the basis of empirically-based reflectance models, and the design and development of improved rendering processes together with a new VR system. Beside this innovative R&D work another key aspect of REALISE is to have computer vision & computer graphics cooperate to produce realistic 3D data efficiently.