Domain Adaptation (DA) has recently been of strong interest in the medical imaging community. While a large variety of DA techniques have been proposed for image segmentation, most of these techniques have been validated either on private datasets or on small publicly available datasets. Moreover, these datasets mostly addressed single-class problems. To tackle these limitations, the Cross-Modality Domain Adaptation (crossMoDA) challenge was organised in conjunction with the 24th International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI 2021). CrossMoDA is the first large and multi-class benchmark for unsupervised cross-modality Domain Adaptation. The goal of the challenge is to segment two key brain structures involved in the follow-up and treatment planning of vestibular schwannoma (VS): the VS and the cochleas. Currently, the diagnosis and surveillance in patients with VS are commonly performed using contrast-enhanced T1 (ceT1) MR imaging. However, there is growing interest in using non-contrast imaging sequences such as high-resolution T2 (hrT2) imaging. For this reason, we established an unsupervised cross-modality segmentation benchmark. The training dataset provides annotated ceT1 scans (N=105) and unpaired non-annotated hrT2 scans (N=105). The aim was to automatically perform unilateral VS and bilateral cochlea segmentation on hrT2 scans as provided in the testing set (N=137). This problem is particularly challenging given the large intensity distribution gap across the modalities and the small volume of the structures. A total of 55 teams from 16 countries submitted predictions to the validation leaderboard. Among them, 16 teams from 9 different countries submitted their algorithm for the evaluation phase. The level of performance reached by the top-performing teams is strikingly high (best median Dice score — VS: 88.4%; Cochleas: 85.7%) and close to full supervision (median Dice score — VS: 92.5%; Cochleas: 87.7%). All top-performing methods made use of an image-to-image translation approach to transform the source-domain images into pseudo-target-domain images. A segmentation network was then trained using these generated images and the manual annotations provided for the source image.

Multi-channel cochlear implants (CI) leverage frequency based cochlear tonotopic mapping to map acoustic information to the cochlear place of stimulation which is primarily determined by electrode locations. Despite the fact that electrode locations within the cochlea are unique to each patient, the acoustic frequencies assigned to the electrodes by the CI processor are determined generically, resulting in a mismatch between intended and actual pitch perception. This is known to be a limiting factor for hearing outcomes with CIs. In this study, we propose a novel, image-guided CI processor programming strategy to select more optimal, patient-customized frequency assignments. The performance of the proposed strategy was evaluated using vocoder-based simulations with ten normal hearing listeners. In our simulations, our strategy results in significantly better speech recognition scores than the standard clinical strategy.

Localizing the anterior and posterior commissures (AC/PC) and the midsagittal plane (MSP) is crucial in stereotactic and functional neurosurgery, human brain mapping, and medical image processing. We present a learning-based method for automatic and efficient localization of these landmarks and the plane using regression forests. Given a point in an image, we first extract a set of multiscale long-range contextual features. We then build random forests models to learn a nonlinear relationship between these features and the probability of the point being a landmark or in the plane. Three-stage coarse-to-fine models are trained for the AC, PC, and MSP separately using downsampled by 4, downsampled by 2, and the original images. Localization is performed hierarchically, starting with a rough estimation that is progressively refined. We evaluate our method using a leave-one-out approach with 100 clinical T1-weighted images and compare it to state-of-the-art methods including an atlas-based approach with six nonrigid registration algorithms and a model-based approach for the AC and PC, and a global symmetry-based approach for the MSP. Our method results in an overall error of 0.55 ±0.30 mm for AC, 0.56 ±0.28 mm for PC, 1.08(°) ±0.66 in the plane's normal direction, and 1.22 ±0.73 voxels in average distance for MSP; it performs significantly better than four registration algorithms and the model-based method for AC and PC, and the global symmetry-based method for MSP. We also evaluate the sensitivity of our method to image quality and parameter values. We show that it is robust to asymmetry, noise, and rotation. Computation time is 25 s.

Identification of error in nonrigid registration is a critical problem in the medical image processing community. We recently proposed an algorithm that we call "Assessing Quality Using Image Registration Circuits" (AQUIRC) to identify nonrigid registration errors and have tested its performance using simulated cases. In this paper, we extend our previous work to assess AQUIRC's ability to detect local nonrigid registration errors and validate it quantitatively at specific clinical landmarks, namely the anterior commissure and the posterior commissure. To test our approach on a representative range of error we utilize five different registration methods and use 100 target images and nine atlas images. Our results show that AQUIRC's measure of registration quality correlates with the true target registration error (TRE) at these selected landmarks with an R(2)=0.542. To compare our method to a more conventional approach, we compute local normalized correlation coefficient (LNCC) and show that AQUIRC performs similarly. However, a multi-linear regression performed with both AQUIRC's measure and LNCC shows a higher correlation with TRE than correlations obtained with either measure alone, thus showing the complementarity of these quality measures. We conclude the paper by showing that the AQUIRC algorithm can be used to reduce registration errors for all five algorithms.

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.

Gradient-echo (GE) echo planar imaging (EPI) is susceptible to both geometric distortions and signal loss. This paper presents a retrospective correction approach based on nonrigid image registration. A new physics-based intensity correction factor derived to compensate for intravoxel dephasing in GE EPI images is incorporated into a previously reported nonrigid registration algorithm. Intravoxel dephasing causes signal loss and thus intensity attenuation in the images. The new rephasing factor we introduce, which changes the intensity of a voxel in images during the registration, is used to improve the accuracy of the intensity-based nonrigid registration method and mitigate the intensity attenuation effect. Simulation-based experiments are first used to evaluate the method. A magnetic resonance (MR) simulator and a real field map are used to generate a realistic GE EPI image. The geometric distortion computed from the field map is used as the ground truth to which the estimated nonrigid deformation is compared. We then apply the algorithm to a set of real human brain images. The results show that, after registration, alignment between EPI and multi-shot, spin-echo images, which have relatively long acquisition times but negligible distortion, is improved and that signal loss caused by dephasing can be recovered.

In current practice, optimal placement of deep-brain stimulators (DBSs) used to treat movement disorders in patients with Parkinson's disease and essential tremor is an iterative procedure. A target is chosen preoperatively based on anatomical landmarks identified on magnetic resonance images. This point is used as an initial position that is refined intraoperatively using both microelectrode recordings and macrostimulation. In this paper, we report on our current progress toward developing a system for the computer-assisted preoperative selection of target points and for the intraoperative adjustment of these points. The system consists of a deformable atlas of optimal target points that can be used to select automatically the preoperative target, of an electrophysiological atlas, and of an intraoperative interface. Results we have obtained show that automatic prediction of target points is an achievable goal. Our results also indicate that electrophysiological information could be used to resolve structures not visible in anatomic images, thus improving both preoperative and intraoperative guidance. Our intraoperative system has reached the stage of a working prototype and we compare targeting accuracy as well as the number of paths needed to reach the targets with our system and with the method in current clinical use.

This paper reports a novel method to track brain shift using a laser-range scanner (LRS) and nonrigid registration techniques. The LRS used in this paper is capable of generating textured point-clouds describing the surface geometry/intensity pattern of the brain as presented during cranial surgery. Using serial LRS acquisitions of the brain's surface and two-dimensional (2-D) nonrigid image registration, we developed a method to track surface motion during neurosurgical procedures. A series of experiments devised to evaluate the performance of the developed shift-tracking protocol are reported. In a controlled, quantitative phantom experiment, the results demonstrate that the surface shift-tracking protocol is capable of resolving shift to an accuracy of approximately 1.6 mm given initial shifts on the order of 15 mm. Furthermore, in a preliminary in vivo case using the tracked LRS and an independent optical measurement system, the automatic protocol was able to reconstruct 50% of the brain shift with an accuracy of 3.7 mm while the manual measurement was able to reconstruct 77% with an accuracy of 2.1 mm. The results suggest that a LRS is an effective tool for tracking brain surface shift during neurosurgery.

Nonrigid registration of medical images is important for a number of applications such as the creation of population averages, atlas-based segmentation, or geometric correction of functional magnetic resonance imaging (fMRI) images to name a few. In recent years, a number of methods have been proposed to solve this problem, one class of which involves maximizing a mutual information (MI)-based objective function over a regular grid of splines. This approach has produced good results but its computational complexity is proportional to the compliance of the transformation required to register the smallest structures in the image. Here, we propose a method that permits the spatial adaptation of the transformation's compliance. This spatial adaptation allows us to reduce the number of degrees of freedom in the overall transformation, thus speeding up the process and improving its convergence properties. To develop this method, we introduce several novelties: 1) we rely on radially symmetric basis functions rather than B-splines traditionally used to model the deformation field; 2) we propose a metric to identify regions that are poorly registered and over which the transformation needs to be improved; 3) we partition the global registration problem into several smaller ones; and 4) we introduce a new constraint scheme that allows us to produce transformations that are topologically correct. We compare the approach we propose to more traditional ones and show that our new algorithm compares favorably to those in current use.

Automated physiologic event detection and alerting is a challenging task in the ICU. Ideally care providers should be alerted only when events are clinically significant and there is opportunity for corrective action. However, the concepts of clinical significance and opportunity are difficult to define in automated systems, and effectiveness of alerting algorithms is difficult to measure. This paper describes recent efforts on the Simon project to capture information from ICU care providers about patient state and therapy in response to alerts, in order to assess the value of event definitions and progressively refine alerting algorithms. Event definitions for intracranial pressure and cerebral perfusion pressure were studied by implementing a reliable system to automatically deliver alerts to clinical users' alphanumeric pagers, and to capture associated documentation about patient state and therapy when the alerts occurred. During a 6-month test period in the trauma ICU at Vanderbilt University Medical Center, 530 alerts were detected in 2280 hours of data spanning 14 patients. Clinical users electronically documented 81% of these alerts as they occurred. Retrospectively classifying documentation based on therapeutic actions taken, or reasons why actions were not taken, provided useful information about ways to potentially improve event definitions and enhance system utility.

This paper describes the definition and testing of a new type of median filter for images. The topological median filter implements some existing ideas and some new ideas on fuzzy connectedness to improve, over a conventional median filter, the extraction of edges in noise. The concept of alpha-connectivity is defined and used to create an algorithm for computing the degree of connectedness of a pixel to all the other pixels in an arbitrary neighborhood. The resulting connectivity map of the neighborhood effectively disconnects peaks in the neighborhood that are separated from the center pixel by a valley in the brightness topology. The median of the connectivity map is an estimate of the median of the peak or plateau to which the center pixel belongs. Unlike the conventional median filter, the topological median is relatively unaffected by disconnected features in the neighborhood of the center pixel. Four topological median filters are defined. Qualitative and statistical analyses of the four filters are presented. It is demonstrated that edge detection can be more accurate on topologically median filtered images than on conventionally median filtered images.

The study presented in this paper tests the hypothesis that the combination of a global similarity transformation and local free-form deformations can be used for the accurate segmentation of internal structures in MR images of the brain. To quantitatively evaluate our approach, the entire brain, the cerebellum, and the head of the caudate have been segmented manually by two raters on one of the volumes (the reference volume) and mapped back onto all the other volumes, using the computed transformations. The contours so obtained have been compared to contours drawn manually around the structures of interest in each individual brain. Manual delineation was performed twice by the same two raters to test inter- and intrarater variability. For the brain and the cerebellum, results indicate that for each rater, contours obtained manually and contours obtained automatically by deforming his own atlas are virtually indistinguishable. Furthermore, contours obtained manually by one rater and contours obtained automatically by deforming this rater's own atlas are more similar than contours obtained manually by two raters. For the caudate, manual intra- and interrater similarity indexes remain slightly better than manual versus automatic indexes, mainly because of the spatial resolution of the images used in this study. Qualitative results also suggest that this method can be used for the segmentation of more complex structures, such as the hippocampus.

Studies aimed at quantifying neuroanatomical differences between populations require the volume measurements of individual brain structures. If the study contains a large number of images, manual segmentation is not practical. This study tests the hypothesis that a fully automatic, atlas-based segmentation method can be used to quantify atrophy indexes derived from the brain and cerebellum volumes in normal subjects and chronic alcoholics. This is accomplished by registering an atlas volume with a subject volume, first using a global transformation, and then improving the registration using a local transformation. Segmented structures in the atlas volume are then mapped to the corresponding structures in the subject volume using the combined global and local transformations. This technique has been applied to seven normal and seven alcoholic subjects. Three magnetic resonance volumes were obtained for each subject and each volume was segmented automatically, using the atlas-based method. Accuracy was assessed by manually segmenting regions and measuring the similarity between corresponding regions obtained automatically. Repeatability was determined by comparing volume measurements of segmented structures from each acquisition of the same subject. Results demonstrate that the method is accurate, that the results are repeatable, and that it can provide a method for automatic quantification of brain atrophy, even when the degree of atrophy is large.

The primary objective of this study is to perform a blinded evaluation of two groups of retrospective image registration techniques, using as a gold standard a prospective marker-based registration method, and to compare the performance of one group with the other. These techniques have already been evaluated individually [27]. In this paper, however, we find that by grouping the techniques as volume based or surface based, we can make some interesting conclusions which were not visible in the earlier study. In order to ensure blindness, all retrospective registrations were performed by participants who had no knowledge of the gold-standard results until after their results had been submitted. Image volumes of three modalities: X-ray computed tomography (CT), magnetic resonance (MR), and positron emission tomography (PET) were obtained from patients undergoing neurosurgery at Vanderbilt University Medical Center on whom bone-implanted fiducial markers were mounted. These volumes had all traces of the markers removed and were provided via the Internet to project collaborators outside Vanderbilt, who then performed retrospective registrations on the volumes, calculating transformations from CT to MR and/or from PET to MR. These investigators communicated their transformations, again via the Internet, to Vanderbilt, where the accuracy of each registration was evaluated. In this evaluation, the accuracy is measured at multiple volumes of interest (VOI's). Our results indicate that the volume-based techniques in this study tended to give substantially more accurate and reliable results than the surface-based ones for the CT-to-MR registration tasks, and slightly more accurate results for the PET-to-MR tasks. Analysis of these results revealed that the rotational component of error was more pronounced for the surface-based group. It was also apparent that all of the registration techniques we examined have the potential to produce satisfactory results much of the time, but that visual inspection is necessary to guard against large errors.

This paper presents a method designed to register preoperative computed tomography (CT) images to vertebral surface points acquired intraoperatively from ultrasound (US) images or via a tracked probe. It also presents a comparison of the registration accuracy achievable with surface points acquired from the entire posterior surface of the vertebra to the accuracy achievable with points acquired only from the spinous process and central laminar regions. Using a marker-based method as a reference, this work shows that submillimetric registration accuracy can be obtained even when a small portion of the posterior vertebral surface is used for registration. It also shows that when selected surface patches are used, CT slice thickness is not a critical parameter in the registration process. Furthermore, the paper includes qualitative results of registering vertebral surface points in US images to multiple CT slices. The method has been tested with US points and physical points on a plastic spine phantom and with simulated data on a patient CT scan.

The authors present a weighted geometrical feature (WGF) registration algorithm. Its efficacy is demonstrated by combining points and a surface. The technique is an extension of Besl and McKay's (1992) iterative closest point (ICP) algorithm. The authors use the WGF algorithm to register X-ray computed tomography (CT) and T2-weighted magnetic resonance (MR) volume head images acquired from eleven patients that underwent craniotomies in a neurosurgical clinical trial. Each patient had five external markers attached to transcutaneous posts screwed into the outer table of the skull. The authors define registration error as the distance between positions of corresponding markers that are not used for registration. The CT and MR images are registered using fiducial paints (marker positions) only, a surface only, and various weighted combinations of points and a surface. The CT surface is derived from contours corresponding to the inner surface of the skull. The MR surface is derived from contours corresponding to the cerebrospinal fluid (CSF)-dura interface. Registration using points and a surface is found to be significantly more accurate then registration using only points or a surface.

It is important to understand any process that affects medical data. Once the data have changed from the original form, one must consider the possibility that the information contained in the data has also changed. In general, false negative and false positive diagnoses caused by this post-processing must be minimized. Medical imaging is one area in which post-processing is commonly performed, but there is often little or no discussion of how these algorithms affect the data. This study uncovers some interesting properties of maximum intensity projection (MIP) algorithms which are commonly used in the post-processing of magnetic resonance (MR) and computed tomography (CT) angiographic data. The appearance of the width of vessels and the extent of malformations such as aneurysms is of interest to clinicians. This study will show how MIP algorithms interact with the shape of the object being projected. MIP's can make objects appear thinner in the projection than in the original data set and also alter the shape of the profile of the object seen in the original data. These effects have consequences for width-measuring algorithms which will be discussed. Each projected intensity is dependent upon the pathlength of the ray from which the projected pixel arises. The morphology (shape and intensity profile) of an object will change the pathlength that each ray experiences. This is termed the pathlength effect. In order to demonstrate the pathlength effect, simple computer models of an imaged vessel were created. Additionally, a static MR phantom verified that the derived equation for the projection-plane probability density function (pdf) predicts the projection-plane intensities well (R(2)=0.96). Finally, examples of projections through in vivo MR angiography and CT angiography data are presented.

The analysis of MR images is evolving from qualitative to quantitative. More and more, the question asked by clinicians is how much and where, rather than a simple statement on the presence or absence of abnormalities. The authors present a study in which the results obtained with a semiautomatic, multispectral segmentation technique are quantitatively compared to manually delineated regions. The core of the semiautomatic image analysis system is a supervised artificial neural network classifier augmented with dedicated preand postprocessing algorithms, including anisotropic noise filtering and a surface-fitting method for the correction of spatial intensity variations. The study was focused on the quantitation of white matter lesions in the human brain. A total of 36 images from six brain volumes was analyzed twice by each of two operators, under supervision of a neuroradiologist. Both the intra- and interrater variability of the methods were studied in terms of the average tissue area detected per slice, the correlation coefficients between area measurements, and a measure of similarity derived from the kappa statistic. The results indicate that, compared to a manual method, the use of the semiautomatic technique not only facilitates the analysis of the images, but also has similar or lower intra- and interrater variabilities.

A number of supervised and unsupervised pattern recognition techniques have been proposed in recent years for the segmentation and the quantitative analysis of MR images. However, the efficacy of these techniques is affected by acquisition artifacts such as inter-slice, intra-slice, and inter-patient intensity variations. Here a new approach to the correction of intra-slice intensity variations is presented. Results demonstrate that the correction process enhances the performance of backpropagation neural network classifiers designed for the segmentation of the images. Two slightly different versions of the method are presented. The first version fits an intensity correction surface directly to reference points selected by the user in the images. The second version fits the surface to reference points obtained by an intermediate classification operation. Qualitative and quantitative evaluation of both methods reveals that the first one leads to a better correction of the images than the second but that it is more sensitive to operator errors.

This work presents an investigation of the potential of artificial neural networks for classification of registered magnetic resonance and X-ray computer tomography images of the human brain. First, topological and learning parameters are established experimentally. Second, the learning and generalization properties of the neural networks are compared to those of a classical maximum likelihood classifier and the superiority of the neural network approach is demonstrated when small training sets are utilized. Third, the generalization properties of the neural networks are utilized to develop an adaptive learning scheme able to overcome interslice intensity variations typical of MR images. This approach permits the segmentation of image volumes based on training sets selected on a single slice. Finally, the segmentation results obtained both with the artificial neural network and the maximum likelihood classifiers are compared to contours drawn manually.

The morphometric analysis of various normal and pathological brain structures is the focus of interest in the study of a number of neuropsychiatric disorders such as Alzheimer's disease (AD) and multi infarct dementia (MID). These studies require the computation of the relative volumes and areas of cerebrospinal fluid (CSF), white matter, gray matter, and white matter lesions (WML) of several types, which are seen in magnetic resonance (MR) images. Such quantitative analyses are facilitated by the extraction of the intracranial cavity from the images, a process that can be demanding and error prone when performed manually. In this paper a robust method is presented that permits its automatic extraction from transverse T2-weighted MR images. The algorithm has been tested on a total of 76 images and the results are discussed.>

A framework for a general signal interpretation system is presented. The structure allows for a collaboration between domain-dependent knowledge (acquired by human experts through years of experience in a specific area) and signal-analysis knowledge (the expertise needed to select and use signal-processing techniques for automated extraction of meaningful features). The system has been built following an object-oriented approach and is organized around a blackboard. Control is handled by a global request-centered mechanism. Morphological and contextual (e.g spatial-temporal relationships) information about the events to be detected by the system is represented in terms of frames. A model interpreter is in charge of matching the events' attributes with the data. A first level of analysis is performed to detect possible candidates. Additional features are opportunistically retrieved from the signal whenever needed by the model interpreter. This permits a focusing on segments of data that are of significance to the reasoning process. The models and the model interpreter have been designed such that the depth of analysis can be adapted to the situation at hand (obvious events will require less feature extraction and computation than doubtful cases). Features needed by the model interpreter are extracted from the signal by independent specialists, which consist of a set of specific digital signal-processing routines and the knowledge required to select the one most appropriate for a given task. The system allows for extensive use of contextual information and adaptation of the event models on the basis of earlier detections. The system has been tested on the problem of automatic electroencephalogram interpretation.>

The design, implementation, and preliminary testing of a computer system for automatic multispectral magnetic resonance imaging analysis is presented. The modular structure of the system permits easy comparison between various classification algorithms. The classification accuracy of traditional statistical pattern-recognition algorithms is compared to the results that can be obtained with neural networks of different topologies. Quantitative (confusion matrices) as well as visual (segmented images) results of a study performed on sets of normal and pathological images are presented. Images segmented with a neural network classifier (NNC) appear less noisy than images segmented with a maximum likelihood classifier (MLC), and it has been observed that the NNC is less sensitive to the selection of the training sets than the MLC

This article describes NetGraph, an object-oriented toolset for risk assessment applications. There are two significant aspects of NetGraph functionality. Firstly, NetGraph is a graphical knowledge acquisition tool for risk assessment, designed for robust structuring and efficient use of domain knowledge. A component library facility in NetGraph helps the domain expert to store and easily retrieve frequently encountered objects in complex systems. Time spent in designing fault trees is significantly reduced by the help of a cut/paste facility that allows the expert to easily define new fault trees by retrieving subtrees (components) of previously defined fault trees and putting them together. A second aspect of NetGraph is fault tree analysis. Fault trees designed in NetGraph can be used by either experts or operators to analyze fault behavior and probability of complex systems. A hypertext interface is also being developed to provide efficient communication of risks to individuals.