• Demonstration of effectiveness of SSL for surgical computer vision using the largest dataset reported to date. • Strong generalization and robust evaluation are shown across six surgical datasets, four procedures, and three tasks, outperforming current SOTA foundation models. • Providing insights into large-scale SSL for surgical computer vision in terms of scaling, pretraining time, dataset composition, and model architecture. • Release of the models and a curated dataset of 2.1 million surgical video frames, establishing a critical resource for advancing surgical foundation model training Foundation models have revolutionized computer vision by achieving vastly superior performance across diverse tasks through large-scale pretraining on extensive datasets. However, their application in surgical computer vision has been limited. This study addresses this gap by introducing SurgeNetXL, a novel surgical foundation model that sets a new benchmark in surgical computer vision. Trained on the largest reported surgical dataset to date, comprising over 4.7 million video frames, SurgeNetXL achieves consistent top-tier performance across six datasets spanning four surgical procedures and three tasks, including semantic segmentation, surgical phase recognition, and critical view of safety (CVS) classification. Compared with the best-performing surgical foundation model, SurgeNetXL shows mean improvements of 4.0%, 8.9%, and 11.4% for semantic segmentation, phase recognition, and CVS classification, respectively. Additionally, SurgeNetXL outperforms ImageNet1k by 16.1%, 8.0%, and 4.3% for the respective tasks. In addition to advancing model performance, this study provides key insights into scaling pretraining datasets, extending training durations, and optimizing model architectures specifically for surgical computer vision. These findings pave the way for improved generalization and robustness in data-scarce scenarios, offering a comprehensive framework for future research in this domain. All models and a subset of the SurgeNetXL dataset, including over 2 million video frames, are publicly available at: https://github.com/TimJaspers0801/SurgeNet .

We introduce LYSTO, the Lymphocyte Assessment Hackathon, which was held in conjunction with the MICCAI 2019 Conference in Shenzhen (China). The competition required participants to automatically assess the number of lymphocytes, in particular T-cells, in images of colon, breast, and prostate cancer stained with CD3 and CD8 immunohistochemistry. Differently from other challenges setup in medical image analysis, LYSTO participants were solely given a few hours to address this problem. In this paper, we describe the goal and the multi-phase organization of the hackathon; we describe the proposed methods and the on-site results. Additionally, we present post-competition results where we show how the presented methods perform on an independent set of lung cancer slides, which was not part of the initial competition, as well as a comparison on lymphocyte assessment between presented methods and a panel of pathologists. We show that some of the participants were capable to achieve pathologist-level performance at lymphocyte assessment. After the hackathon, LYSTO was left as a lightweight plug-and-play benchmark dataset on grand-challenge website, together with an automatic evaluation platform.

Deep learning-based segmentation methods provide an effective and automated way for assessing the structure and function of the heart in cardiac magnetic resonance (CMR) images. However, despite their state-of-the-art performance on images acquired from the same source (same scanner or scanner vendor) as images used during training, their performance degrades significantly on images coming from different domains. A straightforward approach to tackle this issue consists of acquiring large quantities of multi-site and multi-vendor data, which is practically infeasible. Generative adversarial networks (GANs) for image synthesis present a promising solution for tackling data limitations in medical imaging and addressing the generalization capability of segmentation models. In this work, we explore the usability of synthesized short-axis CMR images generated using a segmentation-informed conditional GAN, to improve the robustness of heart cavity segmentation models in a variety of different settings. The GAN is trained on paired real images and corresponding segmentation maps belonging to both the heart and the surrounding tissue, reinforcing the synthesis of semantically-consistent and realistic images. First, we evaluate the segmentation performance of a model trained solely with synthetic data and show that it only slightly underperforms compared to the baseline trained with real data. By further combining real with synthetic data during training, we observe a substantial improvement in segmentation performance (up to 4% and 40% in terms of Dice score and Hausdorff distance) across multiple data-sets collected from various sites and scanner. This is additionally demonstrated across state-of-the-art 2D and 3D segmentation networks, whereby the obtained results demonstrate the potential of the proposed method in tackling the presence of the domain shift in medical data. Finally, we thoroughly analyze the quality of synthetic data and its ability to replace real MR images during training, as well as provide an insight into important aspects of utilizing synthetic images for segmentation.

One of the limiting factors for the development and adoption of novel deep-learning (DL) based medical image analysis methods is the scarcity of labeled medical images. Medical image simulation and synthesis can provide solutions by generating ample training data with corresponding ground truth labels. Despite recent advances, generated images demonstrate limited realism and diversity. In this work, we develop a flexible framework for simulating cardiac magnetic resonance (MR) images with variable anatomical and imaging characteristics for the purpose of creating a diversified virtual population. We advance previous works on both cardiac MR image simulation and anatomical modeling to increase the realism in terms of both image appearance and underlying anatomy. To diversify the generated images, we define parameters: 1)to alter the anatomy, 2) to assign MR tissue properties to various tissue types, and 3) to manipulate the image contrast via acquisition parameters. The proposed framework is optimized to generate a substantial number of cardiac MR images with ground truth labels suitable for downstream supervised tasks. A database of virtual subjects is simulated and its usefulness for aiding a DL segmentation method is evaluated. Our experiments show that training completely with simulated images can perform comparable with a model trained with real images for heart cavity segmentation in mid-ventricular slices. Moreover, such data can be used in addition to classical augmentation for boosting the performance when training data is limited, particularly by increasing the contrast and anatomical variation, leading to better regularization and generalization. The database is publicly available at https://osf.io/bkzhm/ and the simulation code will be available at https://github.com/sinaamirrajab/CMRI.

Accurate segmentation of tubular, network-like structures, such as vessels, neurons, or roads, is relevant to many fields of research. For such structures, the topology is their most important characteristic; particularly preserving connectedness: in the case of vascular networks, missing a connected vessel entirely alters the blood-flow dynamics. We introduce a novel similarity measure termed centerlineDice (short clDice), which is calculated on the inter-section of the segmentation masks and their (morphological) skeleta. We theoretically prove that clDice guarantees topology preservation up to homotopy equivalence for binary 2D and 3D segmentation. Extending this, we pro-pose a computationally efficient, differentiable loss function (soft-clDice) for training arbitrary neural segmentation networks. We benchmark the soft-clDice loss on five public datasets, including vessels, roads and neurons (2D and 3D). Training on soft-clDice leads to segmentation with more accurate connectivity information, higher graph similarity, and better volumetric scores.

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

Deep learning-based methods for deformable image registration are attractive alternatives to conventional registration methods because of their short registration times. However, these methods often fail to estimate larger displacements in complex deformation fields, for which a multi-resolution strategy is required. In this article, we propose to train neural networks progressively to address this problem. Instead of training a large convolutional neural network on the registration task all at once, we initially train smaller versions of the network on lower resolution versions of the images and deformation fields. During training, we progressively expand the network with additional layers that are trained on higher resolution data. We show that this way of training allows a network to learn larger displacements without sacrificing registration accuracy and that the resulting network is less sensitive to large misregistrations compared to training the full network all at once. We generate a large number of ground truth example data by applying random synthetic transformations to a training set of images, and test the network on the problem of intrapatient lung CT registration. We analyze the learned representations in the progressively growing network to assess how the progressive learning strategy influences training. Finally, we show that a progressive training procedure leads to improved registration accuracy when learning large and complex deformations.

Deformable image registration can be time consuming and often needs extensive parameterization to perform well on a specific application. We present a deformable registration method based on a 3-D convolutional neural network, together with a framework for training such a network. The network directly learns transformations between pairs of 3-D images. The network is trained on synthetic random transformations which are applied to a small set of representative images for the desired application. Training, therefore, does not require manually annotated ground truth information on the deformation. The framework for the generation of transformations for training uses a sequence of multiple transformations at different scales that are applied to the image. This way, complex transformations with large displacements can be modeled without folding or tearing images. The methodology is demonstrated on public data sets of inhale-exhale lung CT image pairs which come with landmarks for evaluation of the registration quality. We show that a small training set can be used to train the network, while still allowing generalization to a separate pulmonary CT data set containing data from a different patient group, acquired using a different scanner and scan protocol. This approach results in an accurate and very fast deformable registration method, without a requirement for parameterization at test time or manually annotated data for training.

Accurate segmentation of infant brain magnetic resonance (MR) images into white matter (WM), gray matter (GM), and cerebrospinal fluid (CSF) is an indispensable foundation for early studying of brain growth patterns and morphological changes in neurodevelopmental disorders. Nevertheless, in the isointense phase (approximately 6-9 months of age), due to inherent myelination and maturation process, WM and GM exhibit similar levels of intensity in both T1-weighted (T1w) and T2-weighted (T2w) MR images, making tissue segmentation very challenging. Despite many efforts were devoted to brain segmentation, only few studies have focused on the segmentation of 6-month infant brain images. With the idea of boosting methodological development in the community, iSeg-2017 challenge (http://iseg2017.web.unc.edu) provides a set of 6-month infant subjects with manual labels for training and testing the participating methods. Among the 21 automatic segmentation methods participating in iSeg-2017, we review the 8 top-ranked teams, in terms of Dice ratio, modified Hausdorff distance and average surface distance, and introduce their pipelines, implementations, as well as source codes. We further discuss limitations and possible future directions. We hope the dataset in iSeg-2017 and this review article could provide insights into methodological development for the community.

An unsolved problem in medical image analysis is validation of methods. In this paper we will focus on image registration and in particular on nonlinear image registration, which is one of the hardest analysis problems to validate. The paper covers currently used methods of validation, comparative challenges and public datasets, as well as some of our own work in this area.

This paper presents a robust and fully automatic filter-based approach for retinal vessel segmentation. We propose new filters based on 3D rotating frames in so-called orientation scores, which are functions on the Lie-group domain of positions and orientations [Formula: see text]. By means of a wavelet-type transform, a 2D image is lifted to a 3D orientation score, where elongated structures are disentangled into their corresponding orientation planes. In the lifted domain [Formula: see text], vessels are enhanced by means of multi-scale second-order Gaussian derivatives perpendicular to the line structures. More precisely, we use a left-invariant rotating derivative (LID) frame, and a locally adaptive derivative (LAD) frame. The LAD is adaptive to the local line structures and is found by eigensystem analysis of the left-invariant Hessian matrix (computed with the LID). After multi-scale filtering via the LID or LAD in the orientation score domain, the results are projected back to the 2D image plane giving us the enhanced vessels. Then a binary segmentation is obtained through thresholding. The proposed methods are validated on six retinal image datasets with different image types, on which competitive segmentation performances are achieved. In particular, the proposed algorithm of applying the LAD filter on orientation scores (LAD-OS) outperforms most of the state-of-the-art methods. The LAD-OS is capable of dealing with typically difficult cases like crossings, central arterial reflex, closely parallel and tiny vessels. The high computational speed of the proposed methods allows processing of large datasets in a screening setting.

Accurate localization of the prostate and its surrounding tissue is essential in the treatment of prostate cancer. This paper presents a novel approach to fully automatically segment the prostate, including its seminal vesicles, within a few minutes of a magnetic resonance (MR) scan acquired without an endorectal coil. Such MR images are important in external beam radiation therapy, where using an endorectal coil is highly undesirable. The segmentation is obtained using a deformable model that is trained on-the-fly so that it is specific to the patient's scan. This case specific deformable model consists of a patient specific initialized triangulated surface and image feature model that are trained during its initialization. The image feature model is used to deform the initialized surface by template matching image features (via normalized cross-correlation) to the features of the scan. The resulting deformations are regularized over the surface via well established simple surface smoothing algorithms, which is then made anatomically valid via an optimized shape model. Mean and median Dice's similarity coefficients (DSCs) of 0.85 and 0.87 were achieved when segmenting 3T MR clinical scans of 50 patients. The median DSC result was equal to the inter-rater DSC and had a mean absolute surface error of 1.85 mm. The approach is showed to perform well near the apex and seminal vesicles of the prostate.

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.

Medical image registration is an important task in medical image processing. It refers to the process of aligning data sets, possibly from different modalities (e.g., magnetic resonance and computed tomography), different time points (e.g., follow-up scans), and/or different subjects (in case of population studies). A large number of methods for image registration are described in the literature. Unfortunately, there is not one method that works for all applications. We have therefore developed elastix, a publicly available computer program for intensity-based medical image registration. The software consists of a collection of algorithms that are commonly used to solve medical image registration problems. The modular design of elastix allows the user to quickly configure, test, and compare different registration methods for a specific application. The command-line interface enables automated processing of large numbers of data sets, by means of scripting. The usage of elastix for comparing different registration methods is illustrated with three example experiments, in which individual components of the registration method are varied.

In a multi-atlas based segmentation procedure, propagated atlas segmentations must be combined in a label fusion process. Some current methods deal with this problem by using atlas selection to construct an atlas set either prior to or after registration. Other methods estimate the performance of propagated segmentations and use this performance as a weight in the label fusion process. This paper proposes a selective and iterative method for performance level estimation (SIMPLE), which combines both strategies in an iterative procedure. In subsequent iterations the method refines both the estimated performance and the set of selected atlases. For a dataset of 100 MR images of prostate cancer patients, we show that the results of SIMPLE are significantly better than those of several existing methods, including the STAPLE method and variants of weighted majority voting.

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.

We present a stochastic gradient descent optimisation method for image registration with adaptive step size prediction. The method is based on the theoretical work by Plakhov and Cruz (J. Math. Sci. 120(1):964–973, 2004 ). Our main methodological contribution is the derivation of an image-driven mechanism to select proper values for the most important free parameters of the method. The selection mechanism employs general characteristics of the cost functions that commonly occur in intensity-based image registration. Also, the theoretical convergence conditions of the optimisation method are taken into account. The proposed adaptive stochastic gradient descent (ASGD) method is compared to a standard, non-adaptive Robbins-Monro (RM) algorithm. Both ASGD and RM employ a stochastic subsampling technique to accelerate the optimisation process. Registration experiments were performed on 3D CT and MR data of the head, lungs, and prostate, using various similarity measures and transformation models. The results indicate that ASGD is robust to these variations in the registration framework and is less sensitive to the settings of the user-defined parameters than RM. The main disadvantage of RM is the need for a predetermined step size function. The ASGD method provides a solution for that issue.

Radiation therapy for cervical cancer can benefit from image registration in several ways, for example by studying the motion of organs, or by (partially) automating the delineation of the target volume and other structures of interest. In this paper, the registration of cervical data is addressed using mutual information (MI) of not only image intensity, but also features that describe local image structure. Three aspects of the registration are addressed to make this approach feasible. First, instead of relying on a histogram-based estimation of mutual information, which poses problems for a larger number of features, a graph-based implementation of alpha-mutual information (alpha-MI) is employed. Second, the analytical derivative of alpha-MI is derived. This makes it possible to use a stochastic gradient descent method to solve the registration problem, which is substantially faster than nonderivative-based methods. Third, the feature space is reduced by means of a principal component analysis, which also decreases the registration time. The proposed technique is compared to a standard approach, based on the mutual information of image intensity only. Experiments are performed on 93 T2-weighted MR clinical data sets acquired from 19 patients with cervical cancer. Several characteristics of the proposed algorithm are studied on a subset of 19 image pairs (one pair per patient). On the remaining data (36 image pairs, one or two pairs per patient) the median overlap is shown to improve significantly compared to standard MI from 0.85 to 0.86 for the clinical target volume (CTV, p = 2 x 10(-2)), from 0.75 to 0.81 for the bladder (p = 8 x 10(-6)), and from 0.76 to 0.77 for the rectum (p = 2 x 10(-4)). The registration error is improved at important tissue interfaces, such as that of the bladder with the CTV, and the interface of the rectum with the uterus and cervix.

This paper presents a new method to enforce inverse consistency in nonrigid image registration and matching. Conventional approaches assume diffeomorphic transformation, implicitly or explicitly. However, the inherent smoothness constraint discourages discontinuity consideration. We propose a post-processing algorithm that integrates the input forward and backward fields, which are output by existing registration/matching algorithms, to produce more robust results. Given such a pair of input fields, our algorithm alternately refines the fields by tensor belief propagation, and enforces inverse consistency in stochastic sense by generalized total least squares fitting. To show the efficacy of our stochastic inverse consistency approach, we first present results on very noisy fields. We then demonstrate improvement on existing stereo matching where occlusion is naturally handled by localizing violations of inverse consistency. Finally, we propose a novel application on image stitching, where stochastic inverse consistency is employed in structure deformation, in order to seamlessly align overlapping images with severe misalignment in structure and intensity.

A popular technique for nonrigid registration of medical images is based on the maximization of their mutual information, in combination with a deformation field parameterized by cubic B-splines. The coordinate mapping that relates the two images is found using an iterative optimization procedure. This work compares the performance of eight optimization methods: gradient descent (with two different step size selection algorithms), quasi-Newton, nonlinear conjugate gradient, Kiefer-Wolfowitz, simultaneous perturbation, Robbins-Monro, and evolution strategy. Special attention is paid to computation time reduction by using fewer voxels to calculate the cost function and its derivatives. The optimization methods are tested on manually deformed CT images of the heart, on follow-up CT chest scans, and on MR scans of the prostate acquired using a BFFE, T1, and T2 protocol. Registration accuracy is assessed by computing the overlap of segmented edges. Precision and convergence properties are studied by comparing deformation fields. The results show that the Robbins-Monro method is the best choice in most applications. With this approach, the computation time per iteration can be lowered approximately 500 times without affecting the rate of convergence by using a small subset of the image, randomly selected in every iteration, to compute the derivative of the mutual information. From the other methods the quasi-Newton and the nonlinear conjugate gradient method achieve a slightly higher precision, at the price of larger computation times.

A measure for registration of medical images that currently draws much attention is mutual information. The measure originates from information theory, but has been shown to be successful for image registration as well. Information theory, however, offers many more measures that may be suitable for image registration. These all measure the divergence of the joint distribution of the images' grey values from the joint distribution that would have been found had the images been completely independent. This paper compares the performance of mutual information as a registration measure with that of other F-information measures. The measures are applied to rigid registration of positron emission tomography (PET)/magnetic resonance (MR) and MR/computed tomography (CT) images, for 35 and 41 image pairs, respectively. An accurate gold standard transformation is available for the images, based on implanted markers. The registration performance, robustness and accuracy of the measures are studied. Some of the measures are shown to perform poorly on all aspects. The majority of measures produces results similar to those of mutual information. An important finding, however, is that several measures, although slightly more difficult to optimize, can potentially yield significantly more accurate results than mutual information.

In order to demonstrate the growth of the medical image registration field over the past decades, this paper presents the number of journal publications on this topic since 1988 until 2002. In a similar manner, trends in topics within the field of medical image registration are detected. Publications on computed tomography (CT) and magnetic resonance imaging (MRI) are rather constant through the years. Positron emission tomography (PET) and single photon emission computed tomography (SPECT), on the other hand, seem to loose ground to newly emerging functional imaging techniques, such as functional MRI (fMRI) whereas an increase in interest in registration of ultrasound (US) images was observed. Two topics in image registration that are currently considered hot are intraoperative and elastic registration. Although the interest in intraoperative registration strongly increased in the late 1990s, there seems to be a slight relative decrease in recent years. On the other hand, elastic registration has become a popular topic, reaching the highest numbers so far in 2002.

An overview is presented of the medical image processing literature on mutual-information-based registration. The aim of the survey is threefold: an introduction for those new to the field, an overview for those working in the field, and a reference for those searching for literature on a specific application. Methods are classified according to the different aspects of mutual-information-based registration. The main division is in aspects of the methodology and of the application. The part on methodology describes choices made on facets such as preprocessing of images, gray value interpolation, optimization, adaptations to the mutual information measure, and different types of geometrical transformations. The part on applications is a reference of the literature available on different modalities, on interpatient registration and on different anatomical objects. Comparison studies including mutual information are also considered. The paper starts with a description of entropy and mutual information and it closes with a discussion on past achievements and some future challenges.

Mutual information has developed into an accurate measure for rigid and affine monomodality and multimodality image registration. The robustness of the measure is questionable, however. A possible reason for this is the absence of spatial information in the measure. The present paper proposes to include spatial information by combining mutual information with a term based on the image gradient of the images to be registered. The gradient term not only seeks to align locations of high gradient magnitude, but also aims for a similar orientation of the gradients at these locations. Results of combining both standard mutual information as well as a normalized measure are presented for rigid registration of three-dimensional clinical images [magnetic resonance (MR), computed tomography (CT), and positron emission tomography (PET)]. The results indicate that the combined measures yield a better registration function does mutual information or normalized mutual information per se. The registration functions are less sensitive to low sampling resolution, do not contain incorrect global maxima that are sometimes found in the mutual information function, and interpolation-induced local minima can be reduced. These characteristics yield the promise of more robust registration measures. The accuracy of the combined measures is similar to that of mutual information-based methods.