Spinal cord grey matter segmentation challenge

Compositionally Equivariant Representation Learning

Deep learning models often need sufficient supervision (i.e., labelled data) in order to be trained effectively. By contrast, humans can swiftly learn to identify important anatomy in medical images like MRI and CT scans, with minimal guidance. This recognition capability easily generalises to new images from different medical facilities and to new tasks in different settings. This rapid and generalisable learning ability is largely due to the compositional structure of image patterns in the human brain, which are not well represented in current medical models. In this paper, we study the utilisation of compositionality in learning more interpretable and generalisable representations for medical image segmentation. Overall, we propose that the underlying generative factors that are used to generate the medical images satisfy compositional equivariance property, where each factor is compositional (e.g., corresponds to human anatomy) and also equivariant to the task. Hence, a good representation that approximates well the ground truth factor has to be compositionally equivariant. By modelling the compositional representations with learnable von-Mises-Fisher (vMF) kernels, we explore how different design and learning biases can be used to enforce the representations to be more compositionally equivariant under un-, weakly-, and semi-supervised settings. Extensive results show that our methods achieve the best performance over several strong baselines on the task of semi-supervised domain-generalised medical image segmentation. Code will be made publicly available upon acceptance at https://github.com/vios-s.

Automatic segmentation of tumour and organs at risk in 3D MRI for cervical cancer radiation therapy with anatomical variations

Cervical cancer is a common cancer in women globally, with treatment usually involving radiation therapy (RT). Accurate segmentation for the tumour site and organ-at-risks (OARs) could assist in the reduction of treatment side effects and improve treatment planning efficiency. Cervical cancer Magnetic Resonance Imaging (MRI) segmentation is challenging due to a limited amount of training data available and large inter- and intra- patient shape variation for OARs. The proposed Masked-Net consists of a masked encoder within the 3D U-Net to account for the large shape variation within the dataset, with additional dilated layers added to improve segmentation performance. A new loss function was introduced to consider the bounding box loss during training with the proposed Masked-Net. Transfer learning from a male pelvis MRI data with a similar field of view was included. The approaches were compared to the 3D U-Net which was widely used in MRI image segmentation. The data used consisted of 52 volumes obtained from 23 patients with stage IB to IVB cervical cancer across a maximum of 7 weeks of RT with manually contoured labels including the bladder, cervix, gross tumour volume, uterus and rectum. The model was trained and tested with a 5-fold cross validation. Outcomes were evaluated based on the Dice Similarity Coefficients (DSC), the Hausdorff Distance (HD) and the Mean Surface Distance (MSD). The proposed method accounted for the small dataset, large variations in OAR shape and tumour sizes with an average DSC, HD and MSD for all anatomical structures of 0.790, 30.19mm and 3.15mm respectively.

Improved inter-subject alignment of the lumbosacral cord for group-level in vivo gray and white matter assessments: A scan-rescan MRI study at 3T

INTRODUCTION

Magnetic resonance imaging (MRI) enables the investigation of pathological changes in gray and white matter at the lumbosacral enlargement (LSE) and conus medullaris (CM). However, conducting group-level analyses of MRI metrics in the lumbosacral spinal cord is challenging due to variability in CM length, lack of established image-based landmarks, and unknown scan-rescan reliability. This study aimed to improve inter-subject alignment of the lumbosacral cord to facilitate group-level analyses of MRI metrics. Additionally, we evaluated the scan-rescan reliability of MRI-based cross-sectional area (CSA) measurements and diffusion tensor imaging (DTI) metrics.

METHODS

Fifteen participants (10 healthy volunteers and 5 patients with spinal cord injury) underwent axial T2*-weighted and diffusion MRI at 3T. We assessed the reliability of spinal cord and gray matter-based landmarks for inter-subject alignment of the lumbosacral cord, the inter-subject variability of MRI metrics before and after adjusting for the CM length, the intra- and inter-rater reliability of CSA measurements, and the scan-rescan reliability of CSA measurements and DTI metrics.

RESULTS

The slice with the largest gray matter CSA as an LSE landmark exhibited the highest reliability, both within and across raters. Adjusting for the CM length greatly reduced the inter-subject variability of MRI metrics. The intra-rater, inter-rater, and scan-rescan reliability of MRI metrics were the highest at and around the LSE (scan-rescan coefficient of variation <3% for CSA measurements and <7% for DTI metrics within the white matter) and decreased considerably caudal to it.

CONCLUSIONS

To facilitate group-level analyses, we recommend using the slice with the largest gray matter CSA as a reliable LSE landmark, along with an adjustment for the CM length. We also stress the significance of the anatomical location within the lumbosacral cord in relation to the reliability of MRI metrics. The scan-rescan reliability values serve as valuable guides for power and sample size calculations in future longitudinal studies.

Deep learning for unsupervised domain adaptation in medical imaging: Recent advancements and future perspectives

Deep learning has demonstrated remarkable performance across various tasks in medical imaging. However, these approaches primarily focus on supervised learning, assuming that the training and testing data are drawn from the same distribution. Unfortunately, this assumption may not always hold true in practice. To address these issues, unsupervised domain adaptation (UDA) techniques have been developed to transfer knowledge from a labeled domain to a related but unlabeled domain. In recent years, significant advancements have been made in UDA, resulting in a wide range of methodologies, including feature alignment, image translation, self-supervision, and disentangled representation methods, among others. In this paper, we provide a comprehensive literature review of recent deep UDA approaches in medical imaging from a technical perspective. Specifically, we categorize current UDA research in medical imaging into six groups and further divide them into finer subcategories based on the different tasks they perform. We also discuss the respective datasets used in the studies to assess the divergence between the different domains. Finally, we discuss emerging areas and provide insights and discussions on future research directions to conclude this survey.

Learning with limited annotations: A survey on deep semi-supervised learning for medical image segmentation

Medical image segmentation is a fundamental and critical step in many image-guided clinical approaches. Recent success of deep learning-based segmentation methods usually relies on a large amount of labeled data, which is particularly difficult and costly to obtain, especially in the medical imaging domain where only experts can provide reliable and accurate annotations. Semi-supervised learning has emerged as an appealing strategy and been widely applied to medical image segmentation tasks to train deep models with limited annotations. In this paper, we present a comprehensive review of recently proposed semi-supervised learning methods for medical image segmentation and summarize both the technical novelties and empirical results. Furthermore, we analyze and discuss the limitations and several unsolved problems of existing approaches. We hope this review can inspire the research community to explore solutions to this challenge and further advance the field of medical image segmentation.

USE-Evaluator: Performance metrics for medical image segmentation models supervised by uncertain, small or empty reference annotations in neuroimaging

Performance metrics for medical image segmentation models are used to measure the agreement between the reference annotation and the predicted segmentation. Usually, overlap metrics, such as the Dice, are used as a metric to evaluate the performance of these models in order for results to be comparable. However, there is a mismatch between the distributions of cases and the difficulty level of segmentation tasks in public data sets compared to clinical practice. Common metrics used to assess performance fail to capture the impact of this mismatch, particularly when dealing with datasets in clinical settings that involve challenging segmentation tasks, pathologies with low signal, and reference annotations that are uncertain, small, or empty. Limitations of common metrics may result in ineffective machine learning research in designing and optimizing models. To effectively evaluate the clinical value of such models, it is essential to consider factors such as the uncertainty associated with reference annotations, the ability to accurately measure performance regardless of the size of the reference annotation volume, and the classification of cases where reference annotations are empty. We study how uncertain, small, and empty reference annotations influence the value of metrics on a stroke in-house data set regardless of the model. We examine metrics behavior on the predictions of a standard deep learning framework in order to identify suitable metrics in such a setting. We compare our results to the BRATS 2019 and Spinal Cord public data sets. We show how uncertain, small, or empty reference annotations require a rethinking of the evaluation. The evaluation code was released to encourage further analysis of this topic https://github.com/SophieOstmeier/UncertainSmallEmpty.git.

Advanced Structural Magnetic Resonance Imaging of the Spinal Cord: Technical Aspects and Clinical Use

For a long time, technical obstacles have hampered the acquisition of high-resolution images and the development of reliable processing protocols for spinal cord (SC) MRI. Fortunately, this scenario has changed in the past 5-10 years, due to hardware and software improvements. Nowadays, with advanced protocols, SC MRI is considered a useful tool for several inherited and acquired neurologic diseases, not only for diagnosis approach but also for pathophysiological unraveling and as a biomarker for disease monitoring and clinical trials. In this review, we address advanced SC MRI sequences for macrostructural and microstructural evaluation, useful semiautomatic and automatic processing tools and clinical applications on several neurologic conditions such as hereditary cerebellar ataxia, hereditary spastic paraplegia, motor neuron diseases and multiple sclerosis.

Attention-gated U-Net networks for simultaneous axial/sagittal planes segmentation of injured spinal cords

Magnetic resonance imaging is currently the gold standard for the evaluation of spinal cord injuries. Automatic analysis of these injuries is however challenging, as MRI resolutions vary for different planes of analysis and physiological features are often distorted around these injuries. This study proposes a new CNN-based segmentation method in which information is exchanged between two networks analyzing the scans from different planes. Our aim was to develop a robust method for automatic segmentation of the spinal cord in patients having suffered traumatic injuries. The database consisted of 106 sagittal MRI scans from 94 patients with traumatic spinal cord injuries. Our method used an innovative approach where the scans were analyzed in series under the axial and sagittal plane by two different convolutional networks. The results were compared with those of Deepseg 2D from the Spinal Cord Toolbox (SCT), which was taken as state-of-the-art. Comparisons were evaluated using K-Fold cross-validation combined with statistical t-test results on separate test data. Our method achieved significantly better results than Deepseg 2D, with an average Dice coefficient of 0.95 against 0.88 for Deepseg 2D (p <0.001). Other metrics were also used to compare the segmentations, all of which showed significantly better results for our approach. In this study, we introduce a robust method for spinal cord segmentation which is capable of adequately segmenting spinal cords affected by traumatic injuries, improving upon the methods contained in SCT.

Segment anything model for medical image analysis: An experimental study

Training segmentation models for medical images continues to be challenging due to the limited availability of data annotations. Segment Anything Model (SAM) is a foundation model trained on over 1 billion annotations, predominantly for natural images, that is intended to segment user-defined objects of interest in an interactive manner. While the model performance on natural images is impressive, medical image domains pose their own set of challenges. Here, we perform an extensive evaluation of SAM's ability to segment medical images on a collection of 19 medical imaging datasets from various modalities and anatomies. In our experiments, we generated point and box prompts for SAM using a standard method that simulates interactive segmentation. We report the following findings: (1) SAM's performance based on single prompts highly varies depending on the dataset and the task, from IoU=0.1135 for spine MRI to IoU=0.8650 for hip X-ray. (2) Segmentation performance appears to be better for well-circumscribed objects with prompts with less ambiguity such as the segmentation of organs in computed tomography and poorer in various other scenarios such as the segmentation of brain tumors. (3) SAM performs notably better with box prompts than with point prompts. (4) SAM outperforms similar methods RITM, SimpleClick, and FocalClick in almost all single-point prompt settings. (5) When multiple-point prompts are provided iteratively, SAM's performance generally improves only slightly while other methods' performance improves to the level that surpasses SAM's point-based performance. We also provide several illustrations for SAM's performance on all tested datasets, iterative segmentation, and SAM's behavior given prompt ambiguity. We conclude that SAM shows impressive zero-shot segmentation performance for certain medical imaging datasets, but moderate to poor performance for others. SAM has the potential to make a significant impact in automated medical image segmentation in medical imaging, but appropriate care needs to be applied when using it. Code for evaluation SAM is made publicly available at https://github.com/mazurowski-lab/segment-anything-medical-evaluation.

Adversarial co-training for semantic segmentation over medical images

BACKGROUND AND OBJECTIVE

Abundant labeled data drives the model training for better performance, but collecting sufficient labels is still challenging. To alleviate the pressure of label collection, semi-supervised learning merges unlabeled data into training process. However, the joining of unlabeled data (e.g., data from different hospitals with different acquisition parameters) will change the original distribution. Such a distribution shift leads to a perturbation in the training process, potentially leading to a confirmation bias. In this paper, we investigate distribution shift and develop methods to increase the robustness of our models, with the goal of improving performance in semi-supervised semantic segmentation of medical images. We study distribution shift and increase model robustness to it, for improving practical performance in semi-supervised segmentation over medical images.

METHODS

To alleviate the issue of distribution shift, we introduce adversarial training into the co-training process. We simulate perturbations caused by the distribution shift via adversarial perturbations and introduce the adversarial perturbation to attack the supervised training to improve the robustness against the distribution shift. Benefiting from label guidance, supervised training does not collapse under adversarial attacks. For co-training, two sub-models are trained from two views (over two disjoint subsets of the dataset) to extract different kinds of knowledge independently. Co-training outperforms single-model by integrating both views of knowledge to avoid confirmation bias.

RESULTS

For practicality, we conduct extensive experiments on challenging medical datasets. Experimental results show desirable improvements to state-of-the-art counterparts (Yu and Wang, 2019; Peng et al., 2020; Perone et al., 2019). We achieve a DSC score of 87.37% with only 20% of labels on the ACDC dataset, almost same to using 100% of labels. On the SCGM dataset with more distribution shift, we achieve a DSC score of 78.65% with 6.5% of labels, surpassing 10.30% over Peng et al. (2020). Our evaluative results show superior robustness against distribution shifts in medical scenarios.

CONCLUSION

Empirical results show the effectiveness of our work for handling distribution shift in medical scenarios.

Deep Learning-Based Auto-Segmentation of Spinal Cord Internal Structure of Diffusion Tensor Imaging in Cervical Spondylotic Myelopathy

Cervical spondylotic myelopathy (CSM) is a chronic disorder of the spinal cord. ROI-based features on diffusion tensor imaging (DTI) provide additional information about spinal cord status, which would benefit the diagnosis and prognosis of CSM. However, the manual extraction of the DTI-related features on multiple ROIs is time-consuming and laborious. In total, 1159 slices at cervical levels from 89 CSM patients were analyzed, and corresponding fractional anisotropy (FA) maps were calculated. Eight ROIs were drawn, covering both sides of lateral, dorsal, ventral, and gray matter. The UNet model was trained with the proposed heatmap distance loss for auto-segmentation. Mean Dice coefficients on the test dataset for dorsal, lateral, and ventral column and gray matter were 0.69, 0.67, 0.57, 0.54 on the left side and 0.68, 0.67, 0.59, 0.55 on the right side. The ROI-based mean FA value based on segmentation model strongly correlated with the value based on manual drawing. The percentages of the mean absolute error between the two values of multiple ROIs were 0.07, 0.07, 0.11, and 0.08 on the left side and 0.07, 0.1, 0.1, 0.11, and 0.07 on the right side. The proposed segmentation model has the potential to offer a more detailed spinal cord segmentation and would be beneficial for quantifying a more detailed status of the cervical spinal cord.

Longitudinal assessment of cervical spinal cord compartments in multiple sclerosis

BACKGROUND

Although cervical spinal cord (cSC) area is an established biomarker in MS, there is currently a lack of longitudinal assessments of cSC gray and white matter areas.

OBJECTIVE

We conducted an explorative analysis of longitudinal changes of cSC gray and white matter areas in MS patients.

METHODS

65 MS patients (33 relapsing-remitting; 20 secondary progressive and 12 primary progressive) and 20 healthy controls (HC) received clinical and upper cSC MRI assessments over 1.10±0.28 years. cSC compartments were quantified on MRI using the novel averaged magnetization inversion recovery acquisitions sequence (in-plane resolution=0.67 × 0.67mm²), and in-house developed post-processing methods. Patients were stratified regarding clinical progression.

RESULTS

Patients with clinical progression showed faster reduction of cSC areas over time at the level of cSC enlargement (approximate vertebral level C4-C5) compared to stable patients (p<0.05). In addition, when compared to the rostral-cSC (approximate vertebral level C2-C3), a preferential reduction of cSC and white matter areas over time at the level of cSC enlargement (p<0.05 and p<0.01, respectively) was demonstrated only in patients with clinical progression, but not in stable MS patients and HC. Compared to HC, MS patients showed comparable changes over time in all cSC compartments.

CONCLUSIONS

MS patients with clinical disease progression demonstrate subtle signs of a more pronounced tissue loss at the level of cSC enlargement. Future studies should consider larger sample sizes and more extended observation periods.

Fully Automatic Method for Reliable Spinal Cord Compartment Segmentation in Multiple Sclerosis

BACKGROUND AND PURPOSE

Fully automatic quantification methods of spinal cord compartments are needed to study pathologic changes of the spinal cord GM and WM in MS in vivo. We propose a novel method for automatic spinal cord compartment segmentation (SCORE) in patients with MS.

MATERIALS AND METHODS

The cervical spinal cords of 24 patients with MS and 24 sex- and age-matched healthy controls were scanned on a 3T MR imaging system, including an averaged magnetization inversion recovery acquisition sequence. Three experienced raters manually segmented the spinal cord GM and WM, anterior and posterior horns, gray commissure, and MS lesions. Subsequently, manual segmentations were used to train neural segmentation networks of spinal cord compartments with multidimensional gated recurrent units in a 3-fold cross-validation fashion. Total intracranial volumes were quantified using FreeSurfer.

RESULTS

The intra- and intersession reproducibility of SCORE was high in all spinal cord compartments (eg, mean relative SD of GM and WM: ≤ 3.50% and ≤1.47%, respectively) and was better than manual segmentations (all P < .001). The accuracy of SCORE compared with manual segmentations was excellent, both in healthy controls and in patients with MS (Dice similarity coefficients of GM and WM: ≥ 0.84 and ≥0.92, respectively). Patients with MS had lower total WM areas (P < .05), and total anterior horn areas (P < .01 respectively), as measured with SCORE.

CONCLUSIONS

We demonstrate a novel, reliable quantification method for spinal cord tissue segmentation in healthy controls and patients with MS and other neurologic disorders affecting the spinal cord. Patients with MS have reduced areas in specific spinal cord tissue compartments, which may be used as MS biomarkers.

Cervical spinal cord atrophy in amyotrophic lateral sclerosis across disease stages

OBJECTIVE

Spinal cord degeneration is a hallmark of amyotrophic lateral sclerosis. The assessment of gray matter and white matter cervical spinal cord atrophy across clinical stages defined using the King's staging system could advance the understanding of amyotrophic lateral sclerosis progression.

METHODS

We assessed the in vivo spatial pattern of gray and white matter atrophy along cervical spinal cord (C2 to C6 segments) using 2D phase-sensitive inversion recovery imaging in a cohort of 44 amyotrophic lateral sclerosis patients, evaluating its change across the King's stages and the correlation with disability scored by the amyotrophic lateral sclerosis functional rating scale revised (ALSFRS-R) and disease duration. A mathematical model inferring the potential onset of cervical gray matter atrophy was developed.

RESULTS

In amyotrophic lateral sclerosis patients at King's stage 1, significant cervical spinal cord alterations were mainly identified in gray matter, whereas they involved both gray and white matter in patients at King's stage ≥ 2. Gray and white matter areas correlated with clinical disability at all cervical segments. C3-C4 level was the segment showing early gray matter atrophy starting about 7 to 20 months before symptom onset according to our model.

INTERPRETATION

Our findings suggest that cervical spinal cord atrophy spreads from gray to white matter across King's stages in amyotrophic lateral sclerosis, making spinal cord magnetic resonance imaging an in vivo assessment tool to measure the progression of the disease.

CrossMoDA 2021 challenge: Benchmark of cross-modality domain adaptation techniques for vestibular schwannoma and cochlea segmentation

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.

Quantifying Intraparenchymal Hemorrhage after Traumatic Spinal Cord Injury: A Review of Methodology

Intraparenchymal hemorrhage (IPH) after a traumatic injury has been associated with poor neurological outcomes. Although IPH may result from the initial mechanical trauma, the blood and its breakdown products have potentially deleterious effects. Further, the degree of IPH has been correlated with injury severity and the extent of subsequent recovery. Therefore, accurate evaluation and quantification of IPH following traumatic spinal cord injury (SCI) is important to define treatments' effects on IPH progression and secondary neuronal injury. Imaging modalities, such as magnetic resonance imaging (MRI) and ultrasound (US), have been explored by researchers for the detection and quantification of IPH following SCI. Both quantitative and semiquantitative MRI and US measurements have been applied to objectively assess IPH following SCI, but the optimal methods for doing so are not well established. Studies in animal SCI models (rodent and porcine) have explored US and histological techniques in evaluating SCI and have demonstrated the potential to detect and quantify IPH. Newer techniques using machine learning algorithms (such as convolutional neural networks [CNN]) have also been studied to calculate IPH volume and have yielded promising results. Despite long-standing recognition of the potential pathological significance of IPH within the spinal cord, quantifying IPH with MRI or US is a relatively new area of research. Further studies are warranted to investigate their potential use. Here, we review the different and emerging quantitative MRI, US, and histological approaches used to detect and quantify IPH following SCI.

Reliability of multi-parameter mapping (MPM) in the cervical cord: A multi-center multi-vendor quantitative MRI study

MRI based multicenter studies which target neurological pathologies affecting the spinal cord and brain - including spinal cord injury (SCI) - require standardized acquisition protocols and image processing methods. We have optimized and applied a multi-parameter mapping (MPM) protocol that simultaneously covers the brain and the cervical cord within a traveling heads study across six clinical centers (Leutritz et al., 2020). The MPM protocol includes quantitative maps (magnetization transfer saturation (MT), proton density (PD), longitudinal (R1), and effective transverse (R2*) relaxation rates) sensitive to myelination, water content, iron concentration, and morphometric measures, such as cross-sectional cord area. Previously, we assessed the repeatability and reproducibility of the brain MPM data acquired in the five healthy participants who underwent two scan-rescans (Leutritz et al., 2020). This study focuses on the cervical cord MPM data derived from the same acquisitions to determine its repeatability and reproducibility in the cervical cord. MPM matrices of the cervical cord were generated and processed using the hMRI and the spinal cord toolbox. To determine reliability of the cervical MPM data, the intra-site (i.e., scan-rescan) coefficient of variation (CoV), inter-site CoV, and bias within region of interests (C1, C2 and C3 levels) were determined. The range of the mean intra- and inter-site CoV of MT, R1 and PD was between 2.5% and 12%, and between 1.1% and 4.0% for the morphometric measures. In conclusion, the cervical MPM data showed a high repeatability and reproducibility for key imaging biomarkers and hence can be employed as a standardized tool in multi-center studies, including clinical trials.

Advanced spinal cord MRI in multiple sclerosis: Current techniques and future directions

Spinal cord magnetic resonance imaging (MRI) has a central role in multiple sclerosis (MS) clinical practice for diagnosis and disease monitoring. Advanced MRI sequences capable of visualizing and quantifying tissue macro- and microstructure and reflecting different pathological disease processes have been used in MS research; however, the spinal cord remains under-explored, partly due to technical obstacles inherent to imaging this structure. We propose that the study of the spinal cord merits equal ambition in overcoming technical challenges, and that there is much information to be exploited to make valuable contributions to our understanding of MS. We present a narrative review on the latest progress in advanced spinal cord MRI in MS, covering in the first part structural, functional, metabolic and vascular imaging methods. We focus on recent studies of MS and those making significant technical steps, noting the challenges that remain to be addressed and what stands to be gained from such advances. Throughout we also refer to other works that presend more in-depth review on specific themes. In the second part, we present several topics that, in our view, hold particular potential. The need for better imaging of gray matter is discussed. We stress the importance of developing imaging beyond the cervical spinal cord, and explore the use of ultra-high field MRI. Finally, some recommendations are given for future research, from study design to newer developments in analysis, and the need for harmonization of sequences and methods within the field. This review is aimed at researchers and clinicians with an interest in gaining an overview of the current state of advanced MRI research in this field and what is primed to be the future of spinal cord imaging in MS research.

Optimized multi-echo gradient-echo magnetic resonance imaging for gray and white matter segmentation in the lumbosacral cord at 3 T

Atrophy in the spinal cord (SC), gray (GM) and white matter (WM) is typically measured in-vivo by image segmentation on multi-echo gradient-echo magnetic resonance images. The aim of this study was to establish an acquisition and analysis protocol for optimal SC and GM segmentation in the lumbosacral cord at 3 T. Ten healthy volunteers underwent imaging of the lumbosacral cord using a 3D spoiled multi-echo gradient-echo sequence (Siemens FLASH, with 5 echoes and 8 repetitions) on a Siemens Prisma 3 T scanner. Optimal numbers of successive echoes and signal averages were investigated comparing signal-to-noise (SNR) and contrast-to-noise ratio (CNR) values as well as qualitative ratings for segmentability by experts. The combination of 5 successive echoes yielded the highest CNR between WM and cerebrospinal fluid and the highest rating for SC segmentability. The combination of 3 and 4 successive echoes yielded the highest CNR between GM and WM and the highest rating for GM segmentability in the lumbosacral enlargement and conus medullaris, respectively. For segmenting the SC and GM in the same image, we suggest combining 3 successive echoes. For SC or GM segmentation only, we recommend combining 5 or 3 successive echoes, respectively. Six signal averages yielded good contrast for reliable SC and GM segmentation in all subjects. Clinical applications could benefit from these recommendations as they allow for accurate SC and GM segmentation in the lumbosacral cord.

MobileUNetV3—A Combined UNet and MobileNetV3 Architecture for Spinal Cord Gray Matter Segmentation

The inspection of gray matter (GM) tissue of the human spinal cord is a valuable tool for the diagnosis of a wide range of neurological disorders. Thus, the detection and segmentation of GM regions in magnetic resonance images (MRIs) is an important task when studying the spinal cord and its related medical conditions. This work proposes a new method for the segmentation of GM tissue in spinal cord MRIs based on deep convolutional neural network (CNNs) techniques. Our proposed method, called MobileUNetV3, has a UNet-like architecture, with the MobileNetV3 model being used as a pre-trained encoder. MobileNetV3 is light-weight and yields high accuracy compared with many other CNN architectures of similar size. It is composed of a series of blocks, which produce feature maps optimized using residual connections and squeeze-and-excitation modules. We carefully added a set of upsampling layers and skip connections to MobileNetV3 in order to build an effective UNet-like model for image segmentation. To illustrate the capabilities of the proposed method, we tested it on the spinal cord gray matter segmentation challenge dataset and compared it to a number of recent state-of-the-art methods. We obtained results that outperformed seven methods with respect to five evaluation metrics comprising the dice similarity coefficient (0.87), Jaccard index (0.78), sensitivity (87.20%), specificity (99.90%), and precision (87.96%). Based on these highly competitive results, MobileUNetV3 is an effective deep-learning model for the segmentation of GM MRIs in the spinal cord.

Learning disentangled representations in the imaging domain

Disentangled representation learning has been proposed as an approach to learning general representations even in the absence of, or with limited, supervision. A good general representation can be fine-tuned for new target tasks using modest amounts of data, or used directly in unseen domains achieving remarkable performance in the corresponding task. This alleviation of the data and annotation requirements offers tantalising prospects for applications in computer vision and healthcare. In this tutorial paper, we motivate the need for disentangled representations, revisit key concepts, and describe practical building blocks and criteria for learning such representations. We survey applications in medical imaging emphasising choices made in exemplar key works, and then discuss links to computer vision applications. We conclude by presenting limitations, challenges, and opportunities.

Domain Adaptation for Medical Image Analysis: A Survey

Machine learning techniques used in computer-aided medical image analysis usually suffer from the domain shift problem caused by different distributions between source/reference data and target data. As a promising solution, domain adaptation has attracted considerable attention in recent years. The aim of this paper is to survey the recent advances of domain adaptation methods in medical image analysis. We first present the motivation of introducing domain adaptation techniques to tackle domain heterogeneity issues for medical image analysis. Then we provide a review of recent domain adaptation models in various medical image analysis tasks. We categorize the existing methods into shallow and deep models, and each of them is further divided into supervised, semi-supervised and unsupervised methods. We also provide a brief summary of the benchmark medical image datasets that support current domain adaptation research. This survey will enable researchers to gain a better understanding of the current status, challenges and future directions of this energetic research field.

Extent of cord pathology in the lumbosacral enlargement in non-traumatic versus traumatic spinal cord injury

This study compares remote neurodegenerative changes caudal to a cervical injury in degenerative cervical myelopathy (DCM) (i.e., non-traumatic) and incomplete traumatic spinal cord injury (tSCI) patients, using MRI-based tissue area measurements and diffusion tensor imaging (DTI). Eighteen mild to moderate DCM patients with sensory impairments (mJOA score: 16.2±1.9), 14 incomplete tetraplegic tSCI patients (AIS C&D), and 20 healthy controls were recruited. All participants received DTI and T2*-weighted scans in the lumbosacral enlargement (caudal to injury) and at C2/C3 (rostral to injury). MRI readouts included DTI metrics in the white matter (WM) columns and cross-sectional WM and gray matter area. One-way ANOVA with Tukey's post-hoc comparison (p<0.05) was used to assess group differences. In the lumbosacral enlargement, compared to DCM, tSCI patients exhibited decreased fractional anisotropy in the lateral (tSCI vs. DCM, -11.9%, p=0.007) and ventral WM column (-8.0%, p=0.021), and showed trend toward lower values in the dorsal column (-8.9%, p=0.068). At C2/C3, compared to controls, fractional anisotropy was lower in both groups in the dorsal (DCM vs. controls, -7.9%, p=0.024; tSCI vs. controls, -10.0%, p=0.007) and in the lateral column (DCM: -6.2%, p=0.039; tSCI: -13.3%, p<0.001), while tSCI patients had lower fractional anisotropy than DCM patients in the lateral column (-7.6%, p=0.029). WM areas were not different between patient groups but were lower compared to controls in the lumbosacral enlargement (DCM: -16.9%, p<0.001; tSCI, -10.5%, p=0.043) and at C2/C3 (DCM: -16.0%, p<0.001; tSCI: -18.1%, p<0.001). In conclusion, mild to moderate DCM and incomplete tSCI lead to similar degree of degeneration of the dorsal and lateral columns at C2/C3, but tSCI results in more widespread white matter damage in the lumbosacral enlargement. These remote changes are likely to contribute to the patients' impairment and recovery. DTI is a sensitive tool to assess remote pathological changes in DCM and tSCI patients.

Structural brain and spinal cord damage in symptomatic and pre-symptomatic VAPB-related ALS

INTRODUCTION

The VAPB gene is associated with fALS (fALS 8). This disease presents a variable phenotype and no study sought to characterize its neuroanatomical abnormalities until now. This study aims to evaluate structural brain and spinal cord abnormalities in symptomatic and pre-symptomatic VAPB-related ALS.

METHODS

This cohort included 10 presymptomatic and 20 symptomatic carriers of the Pro56Ser VAPB variant as well as 30 matched controls and 20 individuals with sporadic ALS. They underwent detailed clinical evaluation and MRI in a 3 T scanner. Using volumetric T1 sequence, we computed cerebral cortical thickness (FreeSurfer), basal ganglia volumetry (T1 Multi-atlas) and SC morphometry (SpineSeg). DTI was used to assess white matter integrity (DTI Multi-atlas). Groups were compared using a generalized linear model with Bonferroni-corrected p values<0.05. We also plotted VAPB brain expression map using Allen Human Brain Atlas to compare with imaging findings.

RESULTS

Mean age of presymptomatic and symptomatic subjects were 43.2 and 51.9 years, respectively. Most patients had a predominant lower motor neuron phenotype (16/20). Sleep complaints and tremor were the most frequent additional manifestations. Compared to controls, symptomatic subjects had pallidal, brainstem and SC atrophy, whereas presymptomatic only had SC atrophy. This pattern also contrasted with the sALS group that presented motor cortex and corticospinal abnormalities. Brain structural damage and VAPB expression maps were highly overlapping.

CONCLUSION

VAPB-related ALS has a distinctive structural signature that targets the basal ganglia, brainstem and SC, which are regions with high VAPB expression. Neuroanatomical SC changes are evident before clinical onset of the disease.

Longitudinal changes of spinal cord grey and white matter following spinal cord injury

OBJECTIVES

Traumatic and non-traumatic spinal cord injury produce neurodegeneration across the entire neuraxis. However, the spatiotemporal dynamics of spinal cord grey and white matter neurodegeneration above and below the injury is understudied.

METHODS

We acquired longitudinal data from 13 traumatic and 3 non-traumatic spinal cord injury patients (8-8 cervical and thoracic cord injuries) within 1.5 years after injury and 10 healthy controls over the same period. The protocol encompassed structural and diffusion-weighted MRI rostral (C2/C3) and caudal (lumbar enlargement) to the injury level to track tissue-specific neurodegeneration. Regression models assessed group differences in the temporal evolution of tissue-specific changes and associations with clinical outcomes.

RESULTS

At 2 months post-injury, white matter area was decreased by 8.5% and grey matter by 15.9% in the lumbar enlargement, while at C2/C3 only white matter was decreased (-9.7%). Patients had decreased cervical fractional anisotropy (FA: -11.3%) and increased radial diffusivity (+20.5%) in the dorsal column, while FA was lower in the lateral (-10.3%) and ventral columns (-9.7%) of the lumbar enlargement. White matter decreased by 0.34% and 0.35% per month at C2/C3 and lumbar enlargement, respectively, and grey matter decreased at C2/C3 by 0.70% per month.

CONCLUSIONS

This study describes the spatiotemporal dynamics of tissue-specific spinal cord neurodegeneration above and below a spinal cord injury. While above the injury, grey matter atrophy lagged initially behind white matter neurodegeneration, in the lumbar enlargement these processes progressed in parallel. Tracking trajectories of tissue-specific neurodegeneration provides valuable assessment tools for monitoring recovery and treatment effects.

Generic acquisition protocol for quantitative MRI of the spinal cord

Quantitative spinal cord (SC) magnetic resonance imaging (MRI) presents many challenges, including a lack of standardized imaging protocols. Here we present a prospectively harmonized quantitative MRI protocol, which we refer to as the spine generic protocol, for users of 3T MRI systems from the three main manufacturers: GE, Philips and Siemens. The protocol provides guidance for assessing SC macrostructural and microstructural integrity: T1-weighted and T2-weighted imaging for SC cross-sectional area computation, multi-echo gradient echo for gray matter cross-sectional area, and magnetization transfer and diffusion weighted imaging for assessing white matter microstructure. In a companion paper from the same authors, the spine generic protocol was used to acquire data across 42 centers in 260 healthy subjects. The key details of the spine generic protocol are also available in an open-access document that can be found at https://github.com/spine-generic/protocols . The protocol will serve as a starting point for researchers and clinicians implementing new SC imaging initiatives so that, in the future, inclusion of the SC in neuroimaging protocols will be more common. The protocol could be implemented by any trained MR technician or by a researcher/clinician familiar with MRI acquisition.

Automated claustrum segmentation in human brain MRI using deep learning

In the last two decades, neuroscience has produced intriguing evidence for a central role of the claustrum in mammalian forebrain structure and function. However, relatively few in vivo studies of the claustrum exist in humans. A reason for this may be the delicate and sheet‐like structure of the claustrum lying between the insular cortex and the putamen, which makes it not amenable to conventional segmentation methods. Recently, Deep Learning (DL) based approaches have been successfully introduced for automated segmentation of complex, subcortical brain structures. In the following, we present a multi‐view DL‐based approach to segment the claustrum in T1‐weighted MRI scans. We trained and evaluated the proposed method in 181 individuals, using bilateral manual claustrum annotations by an expert neuroradiologist as reference standard. Cross‐validation experiments yielded median volumetric similarity, robust Hausdorff distance, and Dice score of 93.3%, 1.41 mm, and 71.8%, respectively, representing equal or superior segmentation performance compared to human intra‐rater reliability. The leave‐one‐scanner‐out evaluation showed good transferability of the algorithm to images from unseen scanners at slightly inferior performance. Furthermore, we found that DL‐based claustrum segmentation benefits from multi‐view information and requires a sample size of around 75 MRI scans in the training set. We conclude that the developed algorithm allows for robust automated claustrum segmentation and thus yields considerable potential for facilitating MRI‐based research of the human claustrum. The software and models of our method are made publicly available.

Multi-muscle deep learning segmentation to automate the quantification of muscle fat infiltration in cervical spine conditions

Muscle fat infiltration (MFI) has been widely reported across cervical spine disorders. The quantification of MFI requires time-consuming and rater-dependent manual segmentation techniques. A convolutional neural network (CNN) model was trained to segment seven cervical spine muscle groups (left and right muscles segmented separately, 14 muscles total) from Dixon MRI scans (n = 17, 17 scans < 2 weeks post motor vehicle collision (MVC), and 17 scans 12 months post MVC). The CNN MFI measures demonstrated high test reliability and accuracy in an independent testing dataset (n = 18, 9 scans < 2 weeks post MVC, and 9 scans 12 months post MVC). Using the CNN in 84 participants with scans < 2 weeks post MVC (61 females, 23 males, age = 34.2 ± 10.7 years) differences in MFI between the muscle groups and relationships between MFI and sex, age, and body mass index (BMI) were explored. Averaging across all muscles, females had significantly higher MFI than males (p = 0.026). The deep cervical muscles demonstrated significantly greater MFI than the more superficial muscles (p < 0.001), and only MFI within the deep cervical muscles was moderately correlated to age (r > 0.300, p ≤ 0.001). CNN's allow for the accurate and rapid, quantitative assessment of the composition of the architecturally complex muscles traversing the cervical spine. Acknowledging the wider reports of MFI in cervical spine disorders and the time required to manually segment the individual muscles, this CNN may have diagnostic, prognostic, and predictive value in disorders of the cervical spine.

A pipeline to quantify spinal cord atrophy with deep learning: Application to differentiation of MS and NMOSD patients

PURPOSE

Quantitative measurement of various anatomical regions of the brain and spinal cord (SC) in MRI images are used as unique biomarkers to consider progress and effects of demyelinating diseases of the central nervous system. This paper presents a fully-automated image processing pipeline which quantifies the SC volume of MRI images.

METHODS

In the proposed pipeline, after conducting some pre-processing tasks, a deep convolutional network is utilized to segment the spinal cord cross-sectional area (SCCSA) of each slice. After full segmentation, certain extra slices interpolate between each two adjacent slices using the shape-based interpolation method. Then, a 3D model of the SC is reconstructed, and, by counting the voxels of it, the SC volume is calculated. The performance of the proposed method for the SCCSA segmentation is evaluated on 140 MRI images. Subsequently, to demonstrate the application of the proposed pipeline, we study the differentiations of SC atrophy between 38 Multiple Sclerosis (MS) and 25 Neuromyelitis Optica Spectrum Disorder (NMOSD) patients.

RESULTS

The experimental results of the SCCSA segmentation indicate that the proposed method, adapted by Mask R-CNN, presented the most satisfactory result with the average Dice coefficient of 0.96. For this method, statistical metrics including sensitivity, specificity, accuracy, and precision are 97.51%, 99.98%, 99.92%, and 98.04% respectively. Moreover, the t-test result (p-value = 0.00089) verified a significant difference between the SC atrophy of MS and NMOSD patients.

CONCLUSION

The pipeline efficiently quantifies the SC volume of MRI images and can be utilized as an affordable computer-aided tool for diagnostic purposes.

Deep learning for diabetic retinopathy detection and classification based on fundus images: A review

Diabetic Retinopathy is a retina disease caused by diabetes mellitus and it is the leading cause of blindness globally. Early detection and treatment are necessary in order to delay or avoid vision deterioration and vision loss. To that end, many artificial-intelligence-powered methods have been proposed by the research community for the detection and classification of diabetic retinopathy on fundus retina images. This review article provides a thorough analysis of the use of deep learning methods at the various steps of the diabetic retinopathy detection pipeline based on fundus images. We discuss several aspects of that pipeline, ranging from the datasets that are widely used by the research community, the preprocessing techniques employed and how these accelerate and improve the models' performance, to the development of such deep learning models for the diagnosis and grading of the disease as well as the localization of the disease's lesions. We also discuss certain models that have been applied in real clinical settings. Finally, we conclude with some important insights and provide future research directions.

Exploiting Global Structure Information to Improve Medical Image Segmentation

In this paper, we propose a method to enhance the performance of segmentation models for medical images. The method is based on convolutional neural networks that learn the global structure information, which corresponds to anatomical structures in medical images. Specifically, the proposed method is designed to learn the global boundary structures via an autoencoder and constrain a segmentation network through a loss function. In this manner, the segmentation model performs the prediction in the learned anatomical feature space. Unlike previous studies that considered anatomical priors by using a pre-trained autoencoder to train segmentation networks, we propose a single-stage approach in which the segmentation network and autoencoder are jointly learned. To verify the effectiveness of the proposed method, the segmentation performance is evaluated in terms of both the overlap and distance metrics on the lung area and spinal cord segmentation tasks. The experimental results demonstrate that the proposed method can enhance not only the segmentation performance but also the robustness against domain shifts.

Normalization of Spinal Cord Total Cross-Sectional and Gray Matter Areas as Quantified With Radially Sampled Averaged Magnetization Inversion Recovery Acquisitions

Background: MR imaging of the spinal cord (SC) gray matter (GM) at the cervical and lumbar enlargements' level may be particularly informative in lower motor neuron disorders, e. g., spinal muscular atrophy, but also in other neurodegenerative or autoimmune diseases affecting the SC. Radially sampled averaged magnetization inversion recovery acquisition (rAMIRA) is a novel approach to perform SC imaging in clinical settings with favorable contrast and is well-suited for SC GM quantitation. However, before applying rAMIRA in clinical studies, it is important to understand (i) the sources of inter-subject variability of total SC cross-sectional areas (TCA) and GM area (GMA) measurements in healthy subjects and (ii) their relation to age and sex to facilitate the detection of pathology-associated changes. In this study, we aimed to develop normalization strategies for rAMIRA-derived SC metrics using skull and spine-based metrics to reduce anatomical variability.

Methods: Sixty-one healthy subjects (age range 11–93 years, 37.7% women) were investigated with axial two-dimensional rAMIRA imaging at 3T MRI. Cervical and thoracic levels including the level of the cervical (C4/C5) and lumbar enlargements (T_max) were examined. SC T2-weighted sagittal images and high-resolution 3D whole-brain T1-weighted images were acquired. TCA and GMAs were quantified. Anatomical variables with associations of |r| > 0.30 in univariate association with SC areas, and age and sex were used to construct normalization models using backward selection with TCA_C4/C5 as outcome. The effect of the normalization was assessed by % relative standard deviation (RSD) reductions.

Results: Mean inter-individual variability and the SD of the SC area metrics were considerable: TCA_C4/5: 8.1%/9.0; TCA_Tmax: 8.9%/6.5; GMA_C4/C5: 8.6%/2.2; GMA_Tmax: 12.2%/3.8. Normalization based on sex, brain WM volume, and spinal canal area resulted in RSD reductions of 23.7% for TCAs and 12.0% for GM areas at C4/C5. Normalizations based on the area of spinal canal alone resulted in RSD reductions of 10.2% for TCAs and 9.6% for GM areas at C4/C5, respectively.

Discussion: Anatomic inter-individual variability of SC areas is substantial. This study identified effective normalization models for inter-subject variability reduction in TCA and SC GMA in healthy subjects based on rAMIRA imaging.

SoftSeg: Advantages of soft versus binary training for image segmentation

Most image segmentation algorithms are trained on binary masks formulated as a classification task per pixel. However, in applications such as medical imaging, this "black-and-white" approach is too constraining because the contrast between two tissues is often ill-defined, i.e., the voxels located on objects' edges contain a mixture of tissues (a partial volume effect). Consequently, assigning a single "hard" label can result in a detrimental approximation. Instead, a soft prediction containing non-binary values would overcome that limitation. In this study, we introduce SoftSeg, a deep learning training approach that takes advantage of soft ground truth labels, and is not bound to binary predictions. SoftSeg aims at solving a regression instead of a classification problem. This is achieved by using (i) no binarization after preprocessing and data augmentation, (ii) a normalized ReLU final activation layer (instead of sigmoid), and (iii) a regression loss function (instead of the traditional Dice loss). We assess the impact of these three features on three open-source MRI segmentation datasets from the spinal cord gray matter, the multiple sclerosis brain lesion, and the multimodal brain tumor segmentation challenges. Across multiple random dataset splittings, SoftSeg outperformed the conventional approach, leading to an increase in Dice score of 2.0% on the gray matter dataset (p=0.001), 3.3% for the brain lesions, and 6.5% for the brain tumors. SoftSeg produces consistent soft predictions at tissues' interfaces and shows an increased sensitivity for small objects (e.g., multiple sclerosis lesions). The richness of soft labels could represent the inter-expert variability, the partial volume effect, and complement the model uncertainty estimation, which is typically unclear with binary predictions. The developed training pipeline can easily be incorporated into most of the existing deep learning architectures. SoftSeg is implemented in the freely-available deep learning toolbox ivadomed (https://ivadomed.org).

Spinal Cord Gray and White Matter Damage in Different Hereditary Spastic Paraplegia Subtypes

BACKGROUND AND PURPOSE

Spinal cord damage is a hallmark of hereditary spastic paraplegias, but it is still not clear whether specific subtypes of the disease have distinctive patterns of spinal cord gray (GM) and white (WM) matter involvement. We compared cervical cross-sectional GM and WM areas in patients with distinct hereditary spastic paraplegia subtypes. We also assessed whether these metrics correlated with clinical parameters.

MATERIALS AND METHODS

We analyzed 37 patients (17 men; mean age, 47.3 [SD, 16.5] years) and 21 healthy controls (7 men; mean age, 42.3 [SD, 13.2] years). There were 7 patients with spastic paraplegia type 3A (SPG3A), 12 with SPG4, 10 with SPG7, and 8 with SPG11. Image acquisition was performed on a 3T MR imaging scanner, and T2*-weighted 2D images were assessed by the Spinal Cord Toolbox. Statistical analyses were performed in SPSS using nonparametric tests and false discovery rate-corrected P values < .05.

RESULTS

The mean disease duration for the hereditary spastic paraplegia group was 22.4 [SD, 13.8] years and the mean Spastic Paraplegia Rating Scale score was 22.8 [SD, 11.0]. We failed to identify spinal cord atrophy in SPG3A and SPG7. In contrast, we found abnormalities in patients with SPG4 and SPG11. Both subtypes had spinal cord GM and WM atrophy. SPG4 showed a strong inverse correlation between GM area and disease duration (ρ = -0.903, P < .001).

CONCLUSIONS

Cervical spinal cord atrophy is found in some but not all hereditary spastic paraplegia subtypes. Spinal cord damage in SPG4 and 11 involves both GM and WM.

Magnetization transfer and diffusion tensor imaging in dogs with intervertebral disk herniation

Background

Quantitative magnetic resonance imaging (QMRI) techniques of magnetization transfer ratio (MTR) and diffusion tensor imaging (DTI) provide microstructural information about the spinal cord.

Objective

Compare neurologic grades using the modified Frankel scale with MTR and DTI measurements in dogs with thoracolumbar intervertebral disk herniation (IVDH).

Animals

Fifty‐one dogs with thoracolumbar IVDH.

Methods

Prospective cohort study. Quantitative MRI measurements of the spinal cord were obtained at the region of compression. A linear regression generalized estimating equations model was used to compare QMRI measurements between different neurological grades after adjusting for age, weight, duration of clinical signs, and lesion location.

Results

Grade 5 (.79  ×  10⁻³ mm²/s [median], .43−.91 [range]) and axial (1.47 × 10⁻³ mm²/s, .58−1.8) diffusivity were lower compared to grades 2 (1.003, .68−1.36; P = .02 and 1.81 × 10⁻³ mm²/s, 1.36−2.12; P < .001, respectively) and 3 (1.07 × 10⁻³ mm²/s, .77−1.5; P = .04 and 1.92 × 10⁻³ mm²/s, 1.83−2.37;P < .001, respectively). Compared to dogs with acute myelopathy, chronic myelopathy was associated with higher mean (1.02 × 10⁻³ mm²/s, .77−1.36 vs. .83 × 10⁻³ mm²/s, .64−1.5; P = .03) and radial diffusivity (.75 × 10⁻³ mm²/s, .38−1.04 vs. .44 × 10⁻³ mm²/s, .22−1.01; P = .008) and lower MTR (46.76, 31.8−56.43 vs. 54.4, 45.2−62.27; P = .004) and fractional anisotropy (.58, .4−0.75 vs. .7, .46−.85; P = .02). Fractional anisotropy was lower in dogs with a T2‐weighted intramedullary hyperintensity compared to those without (.7, .45−.85 vs. .54, .4−.8; P = .01).

Conclusion and Clinical Relevance

Mean diffusivity and AD could serve as surrogates of severity of spinal cord injury and are complementary to the clinical exam in dogs with thoracolumbar IVDH.

Imaging of the spine and spinal cord: An overview of magnetic resonance imaging (MRI) techniques

This review will discuss conventional and advanced magnetic resonance (MRI) imaging techniques used to study the spine and spinal cord according to the anatomical structures and clinical indications. Clinical challenges that neuroradiologists may face are also discussed, such as the "when" and "where" concerning the use of each technique, and in which pathology or clinical scenario each technique is useful. Finally, some "tips and tricks" to overcome the challenges are provided with clinical examples.

Voxelwise analysis of diffusion MRI of cervical spinal cord using tract-based spatial statistics

Robust voxelwise analysis using tract-based spatial statistics (TBSS) together with permutation statistical method is standardly used in analyzing diffusion tensor imaging (DTI) of brain. A similar analytical method could be useful when studying DTI of cervical spinal cord. Based on anatomical data of sixty-four healthy volunteers, white (WM) and gray matter (GM) masks were created and subsequently registered into DTI space. Using TBSS, two skeleton types were created (single line and dilated for WM as well as GM). From anatomical data, percentage rates of overlap were calculated for all skeletons in relation to WM and GM masks. Voxelwise analysis of fractional anisotropy values depending on age and sex was conducted. Correlation of fraction anisotropy values with age of subjects was also evaluated. The two WM skeleton types showed a high overlap rate with WM masks (~94%); GM skeletons showed lower rates (56% and 42%, respectively, for single line and dilated). WM and GM areas where fraction anisotropy values differ between sexes were identified (p < .05). Furthermore, using voxelwise analysis such WM voxels were identified where fraction anisotropy values differ depending on age (p < .05) and in these voxels linear dependence of fraction anisotropy and age (r = -0.57, p < .001) was confirmed by regression analysis. This dependence was not proven when using WM anatomical masks (r = -0.21, p = .10). The analytical approach presented shown to be useful for group analysis of DTI data for cervical spinal cord.

Effects of Deep Learning Reconstruction Technique in High-Resolution Non-contrast Magnetic Resonance Coronary Angiography at a 3-Tesla Machine

Purpose:

To evaluate the effects of deep learning reconstruction (DLR) in qualitative and quantitative image quality of non-contrast magnetic resonance coronary angiography (MRCA).

Methods:

Ten healthy volunteers underwent conventional MRCA (C-MRCA) and high-resolution (HR) MRCA on a 3T magnetic resonance imaging with a voxel size of 1.8 × 1.1 × 1.7 mm3 and 1.8 × 0.6 × 1.0 mm3, respectively, for C-MRCA and HR-MRCA. High-resolution magnetic resonance coronary angiography was also reconstructed with the DLR technique (DLR-HR-MRCA). We compared the contrast-to-noise ratio (CNR) and visual evaluation scores for vessel sharpness and traceability of proximal and distal coronary vessels on a 4-point scale among 3 image series.

Results:

The vascular CNR value on the C-MRCA and the DLR-HR-MRCA was significantly higher than that on the HR-MRCA in the proximal and distal coronary arteries (13.9 ± 6.4, 11.3 ± 4.4, and 7.8 ± 2.6 for C-MRCA, DLR-HR-MRCA, and HR-MRCA, P < .05, respectively). Mean visual evaluation scores for the vessel sharpness and traceability of proximal and distal coronary vessels were significantly higher on the HR-DLR-MRCA than the C-MRCA ( P < .05, respectively).

Conclusion:

Deep learning reconstruction significantly improved the CNR of coronary arteries on HR-MRCA, resulting in both higher visual image quality and better vessel traceability compared with C-MRCA.

Probabilistic Learning Coherent Point Drift for 3D Ultrasound Fetal Head Registration

Quantification of brain growth is crucial for the assessment of fetal well being, for which ultrasound (US) images are the chosen clinical modality. However, they present artefacts, such as acoustic occlusion, especially after the 18^th gestational week, when cranial calcification appears. Fetal US volume registration is useful in one or all of the following cases: to monitor the evolution of fetometry indicators, to segment different structures using a fetal brain atlas, and to align and combine multiple fetal brain acquisitions. This paper presents a new approach for automatic registration of real 3D US fetal brain volumes, volumes that contain a considerable degree of occlusion artefacts, noise, and missing data. To achieve this, a novel variant of the coherent point drift method is proposed. This work employs supervised learning to segment and conform a point cloud automatically and to estimate their subsequent weight factors. These factors are obtained by a random forest-based classification and are used to appropriately assign nonuniform membership probability values of a Gaussian mixture model. These characteristics allow for the automatic registration of 3D US fetal brain volumes with occlusions and multiplicative noise, without needing an initial point cloud. Compared to other intensity and geometry-based algorithms, the proposed method achieves an error reduction of 7.4% to 60.7%, with a target registration error of only 6.38 ± 3.24 mm. This makes the herein proposed approach highly suitable for 3D automatic registration of fetal head US volumes, an approach which can be useful to monitor fetal growth, segment several brain structures, or even compound multiple acquisitions taken from different projections.

Open-source pipeline for multi-class segmentation of the spinal cord with deep learning

This paper presents an open-source pipeline to train neural networks to segment structures of interest from MRI data. The pipeline is tailored towards homogeneous datasets and requires relatively low amounts of manual segmentations (few dozen, or less depending on the homogeneity of the dataset). Two use-case scenarios for segmenting the spinal cord white and grey matter are presented: one in marmosets with variable numbers of lesions, and the other in the publicly available human grey matter segmentation challenge [1]. The pipeline is freely available at: https://github.com/neuropoly/multiclass-segmentation.

Traumatic and nontraumatic spinal cord injury: pathological insights from neuroimaging

Pathophysiological changes in the spinal cord white and grey matter resulting from injury can be observed with MRI techniques. These techniques provide sensitive markers of macrostructural and microstructural tissue integrity, which correlate with histological findings. Spinal cord MRI findings in traumatic spinal cord injury (tSCI) and nontraumatic spinal cord injury - the most common form of which is degenerative cervical myelopathy (DCM) - have provided important insights into the pathophysiological processes taking place not just at the focal injury site but also rostral and caudal to the spinal injury. Although tSCI and DCM have different aetiologies, they show similar degrees of spinal cord pathology remote from the injury site, suggesting the involvement of similar secondary degenerative mechanisms. Advanced quantitative MRI protocols that are sensitive to spinal cord pathology have the potential to improve diagnosis and, more importantly, predict outcomes in patients with tSCI or nontraumatic spinal cord injury. This Review describes the insights into tSCI and DCM that have been revealed by neuroimaging and outlines current activities and future directions for the field.

Intersubject Variability and Normalization Strategies for Spinal Cord Total Cross-Sectional and Gray Matter Areas

BACKGROUND AND PURPOSE

The quantification of spinal cord (SC) atrophy by MRI has assumed an important role in assessment of neuroinflammatory/neurodegenerative diseases and traumatic SC injury. Recent technical advances make possible the quantification of gray matter (GM) and white matter tissues in clinical settings. However, the goal of a reliable diagnostic, prognostic or predictive marker is still elusive, in part due to large intersubject variability of SC areas. Here, we investigated the sources of this variability and explored effective strategies to reduce it.

METHODS

One hundred twenty-nine healthy subjects (mean age: 41.0 ± 15.9) underwent MRI on a Siemens 3T Skyra scanner. Two-dimensional PSIR at the C2-C3 vertebral level and a sagittal 1 mm³ 3D T1-weighted brain acquisition extended to the upper cervical cord were acquired. Total cross-sectional area and GM area were measured at C2-C3, as well as measures of the vertebra, spinal canal and the skull. Correlations between the different metrics were explored using Pearson product-moment coefficients. The most promising metrics were used to normalize cord areas using multiple regression analyses.

RESULTS

The most effective normalization metrics were the V-scale (from SienaX) and the product of the C2-C3 spinal canal diameters. Normalization methods based on these metrics reduced the intersubject variability of cord areas of up to 17.74%. The measured cord areas had a statistically significant sex difference, while the effect of age was moderate.

CONCLUSIONS

The present work explored in a large cohort of healthy subjects the source of intersubject variability of SC areas and proposes effective normalization methods for its reduction.

Automatic Spinal Cord Gray Matter Quantification: A Novel Approach

BACKGROUND AND PURPOSE

Currently, accurate and reproducible spinal cord GM segmentation remains challenging and a noninvasive broadly accepted reference standard for spinal cord GM measurements is still a matter of ongoing discussion. Our aim was to assess the reproducibility and accuracy of cervical spinal cord GM and WM cross-sectional area measurements using averaged magnetization inversion recovery acquisitions images and a fully-automatic postprocessing segmentation algorithm.

MATERIALS AND METHODS

The cervical spinal cord of 24 healthy subjects (14 women; mean age, 40 ± 11 years) was scanned in a test-retest fashion on a 3T MR imaging system. Twelve axial averaged magnetization inversion recovery acquisitions slices were acquired over a 48-mm cord segment. GM and WM were both manually segmented by 2 experienced readers and compared with an automatic variational segmentation algorithm with a shape prior modified for 3D data with a slice similarity prior. Precision and accuracy of the automatic method were evaluated using coefficients of variation and Dice similarity coefficients.

RESULTS

The mean GM area was 17.20 ± 2.28 mm² and the mean WM area was 72.71 ± 7.55 mm² using the automatic method. Reproducibility was high for both methods, while being better for the automatic approach (all mean automatic coefficients of variation, ≤4.77%; all differences, P < .001). The accuracy of the automatic method compared with the manual reference standard was excellent (mean Dice similarity coefficients: 0.86 ± 0.04 for GM and 0.90 ± 0.03 for WM). The automatic approach demonstrated similar coefficients of variation between intra- and intersession reproducibility as well as among all acquired spinal cord slices.

CONCLUSIONS

Our novel approach including the averaged magnetization inversion recovery acquisitions sequence and a fully-automated postprocessing segmentation algorithm demonstrated an accurate and reproducible spinal cord GM and WM segmentation. This pipeline is promising for both the exploration of longitudinal structural GM changes and application in clinical settings in disorders affecting the spinal cord.

Multiple sclerosis lesions affect intrinsic functional connectivity of the spinal cord

Patients with multiple sclerosis present with focal lesions throughout the spinal cord. There is a clinical need for non-invasive measurements of spinal cord activity and functional organization in multiple sclerosis, given the cord's critical role in the disease. Recent reports of spontaneous blood oxygenation level-dependent fluctuations in the spinal cord using functional MRI suggest that, like the brain, cord activity at rest is organized into distinct, synchronized functional networks among grey matter regions, likely related to motor and sensory systems. Previous studies looking at stimulus-evoked activity in the spinal cord of patients with multiple sclerosis have demonstrated increased levels of activation as well as a more bilateral distribution of activity compared to controls. Functional connectivity studies of brain networks in multiple sclerosis have revealed widespread alterations, which may take on a dynamic trajectory over the course of the disease, with compensatory increases in connectivity followed by decreases associated with structural damage. We build upon this literature by examining functional connectivity in the spinal cord of patients with multiple sclerosis. Using ultra-high field 7 T imaging along with processing strategies for robust spinal cord functional MRI and lesion identification, the present study assessed functional connectivity within cervical cord grey matter of patients with relapsing-remitting multiple sclerosis (n = 22) compared to a large sample of healthy controls (n = 56). Patient anatomical images were rated for lesions by three independent raters, with consensus ratings revealing 19 of 22 patients presented with lesions somewhere in the imaged volume. Linear mixed models were used to assess effects of lesion location on functional connectivity. Analysis in control subjects demonstrated a robust pattern of connectivity among ventral grey matter regions as well as a distinct network among dorsal regions. A gender effect was also observed in controls whereby females demonstrated higher ventral network connectivity. Wilcoxon rank-sum tests detected no differences in average connectivity or power of low frequency fluctuations in patients compared to controls. The presence of lesions was, however, associated with local alterations in connectivity with differential effects depending on columnar location. The patient results suggest that spinal cord functional networks are generally intact in relapsing-remitting multiple sclerosis but that lesions are associated with focal abnormalities in intrinsic connectivity. These findings are discussed in light of the current literature on spinal cord functional MRI and the potential neurological underpinnings.

A surrogate urethra for real-time planning of high-dose-rate prostate brachytherapy

PURPOSE

This study characterizes prostatic urethra cross-section to develop a surrogate urethra for accurate prediction of urethral dose during real-time high-dose-rate prostate brachytherapy.

MATERIALS AND METHODS

Archived preoperative transrectal ultrasound images from 100 patients receiving low-dose-rate prostate brachytherapy were used to characterize the prostatic urethra, contoured on ultrasound using aerated gel. Consensus contours, defined using majority vote, described commonalities in cross-sectional shape across patients. Potential simplified surrogates were defined and evaluated against the true urethra. The best performing surrogate, a circle of varying size (CS) was retrospectively contoured on 85 high-dose-rate prostate brachytherapy treatment plans. Dose to this recommended surrogate was compared with urethral doses estimated by the standard 6 mm circle surrogate.

RESULTS

Clear variation in urethral cross-sectional shape was observed along its length and between patients. The standard circle surrogate had low predictive sensitivity (61.1%) compared with true urethra because of underrepresentation of the verumontanum midgland. The CS best represented the true urethra across all validation metrics (dice: 0.73, precision: 67.0%, sensitivity: 83.2%, conformity: 0.78). Retrospective evaluation of planned doses using the CS surrogate resulted in significant differences in all reported urethral dose parameters compared with the standard circle, with the exception of D_100%. The urethral dose limit (115%) was exceeded in 40% of patients for the CS surrogate.

CONCLUSIONS

The proposed CS surrogate, consisting of circles of varying diameter, is simple yet better represents the true urethra compared with the standard 6 mm circle. Higher urethral doses were predicted using CS, and the improved accuracy of CS may offer increased predictive power for urethral toxicity, a subject of future work.