PaperlayerView exact revision on arXiv

Source-linked AI summary

Automatic segmentation of the spinal cord and intramedullary multiple sclerosis lesions with convolutional neural networks

Charley Gros, Benjamin De Leener, Atef Badji, Josefina Maranzano, Dominique Eden, Sara M. Dupont, Jason Talbott, Ren Zhuoquiong, Yaou Liu, Tobias Granberg, Russell Ouellette, Yasuhiko Tachibana, Masaaki Hori, Kouhei Kamiya, Lydia Chougar, Leszek Stawiarz, Jan Hillert, Elise Bannier, Anne Kerbrat, Gilles Edan, Pierre Labauge, Virginie Callot, Jean Pelletier, Bertrand Audoin, Henitsoa Rasoanandrianina, Jean-Christophe Brisset, Paola Valsasina, Maria A. Rocca, Massimo Filippi, Rohit Bakshi, Shahamat Tauhid, Ferran Prados, Marios Yiannakas, Hugh Kearney, Olga Ciccarelli, Seth Smith, Constantina Andrada Treaba, Caterina Mainero, Jennifer Lefeuvre, Daniel S. Reich, Govind Nair, Vincent Auclair, Donald G. McLaren, Allan R. Martin, Michael G. Fehlings, Shahabeddin Vahdat, Ali Khatibi, Julien Doyon, Timothy Shepherd, Erik Charlson, Sridar Narayanan, Julien Cohen-Adad

arXiv:1805.06349v2cs.CV

TL;DR

Robust spinal cord and intramedullary MS lesion segmentation remains difficult across varied spinal cord images and multi-site clinical data. This study introduces a two-CNN framework and reports spinal cord segmentation outperforming PropSeg, while lesion results were generally within the range of manual segmentations.

  • Problem

    Existing spinal cord segmentation methods are often restricted by region, contrast, or resolution, while brain lesion methods do not readily transfer to spinal cord morphology and imaging challenges.

  • Method

    The framework uses two CNNs: one detects the spinal cord centerline, and the other segments the spinal cord or intramedullary MS lesions within a localized volume.

  • Results

    Spinal cord segmentation outperformed a state-of-the-art method on a multi-site heterogeneous clinical dataset, while lesion segmentation was generally within the range of manual segmentations.

  • Takeaways & Limitations

    The open-source automatic methods are readily accessible in the Spinal Cord Toolbox for spinal cord and MS lesion segmentation.

  • Takeaways & Limitations

    Performance is likely best on data acquired with parameters similar to those used for training, with variable detectability across sites and false positives or negatives especially affecting small lesions.

Abstract

from arXiv · show

The spinal cord is frequently affected by atrophy and/or lesions in multiple sclerosis (MS) patients. Segmentation of the spinal cord and lesions from MRI data provides measures of damage, which are key criteria for the diagnosis, prognosis, and longitudinal monitoring in MS. Automating this operation eliminates inter-rater variability and increases the efficiency of large-throughput analysis pipelines. Robust and reliable segmentation across multi-site spinal cord data is challenging because of the large variability related to acquisition parameters and image artifacts. The goal of this study was to develop a fully-automatic framework, robust to variability in both image parameters and clinical condition, for segmentation of the spinal cord and intramedullary MS lesions from conventional MRI data. Scans of 1,042 subjects (459 healthy controls, 471 MS patients, and 112 with other spinal pathologies) were included in this multi-site study (n=30). Data spanned three contrasts (T1-, T2-, and T2*-weighted) for a total of 1,943 volumes. The proposed cord and lesion automatic segmentation approach is based on a sequence of two Convolutional Neural Networks (CNNs). To deal with the very small proportion of spinal cord and/or lesion voxels compared to the rest of the volume, a first CNN with 2D dilated convolutions detects the spinal cord centerline, followed by a second CNN with 3D convolutions that segments the spinal cord and/or lesions. When compared against manual segmentation, our CNN-based approach showed a median Dice of 95% vs. 88% for PropSeg, a state-of-the-art spinal cord segmentation method. Regarding lesion segmentation on MS data, our framework provided a Dice of 60%, a relative volume difference of -15%, and a lesion-wise detection sensitivity and precision of 83% and 77%, respectively. The proposed framework is open-source and readily available in the Spinal Cord Toolbox.

4 McConnell Brain Imaging Centre, Montreal Neurological Institute, Montreal, Canada

The listed affiliations include radiology departments and neuroscience research units in China, France, and Italy.

  • Radiology affiliations are listed for Xuanwu Hospital and Beijing Tiantan Hospital in Beijing, China.
  • The French affiliation includes the Observatoire Français de la Sclérose en Plaques and Lyon-based academic and clinical institutions.
  • The Italian affiliation is the Neuroimaging Research Unit at San Raffaele Scientific Institute and Vita-Salute San Raffaele University in Milan.

18 Brigham and Women's Hospital, Harvard Medical School, Boston, USA

The listed affiliations include institutions in London, Maryland, and Toronto associated with neurology, medical imaging, and neurosurgery.

  • The Queen Square MS Centre is affiliated with University College London in London, UK.
  • The National Institute of Neurological Disorders and Stroke is listed in Maryland, USA.
  • The Center for Medical Image Computing is affiliated with University College London in London, United Kingdom.
  • A Division of Neurosurgery affiliation is listed at the University of Toronto in Toronto, Canada.

30 NYU Langone Medical Center, New York, USA

The paper concerns convolutional-neural-network-based spinal-cord and multiple-sclerosis lesion segmentation using MRI data and the Spinal Cord Toolbox.

  • The study uses MRI segmentation for the spinal cord and multiple sclerosis lesions.
  • The method is based on convolutional neural networks, abbreviated CNNs.
  • The data comprise 1,943 volumes spanning T1-, T2-, and T2*-weighted contrasts.
  • SCT denotes the Spinal Cord Toolbox.

1. Introduction

The introduction motivates automatic spinal-cord and intramedullary MS-lesion segmentation by clinical relevance and the difficulty of robust multi-site MRI analysis. The proposed framework uses CNNs and is evaluated across heterogeneous clinical imaging data.

  • Clinical motivation: Spinal-cord lesions and atrophy are clinically relevant because MRI quantification supports MS diagnosis, prognosis, monitoring, and disability-related assessment.
  • Segmentation challenge: Existing spinal-cord segmentation methods may depend on specific regions of interest, contrasts, or resolutions and lack broad multi-site validation.
  • Segmentation challenge: Spinal-cord lesion segmentation is difficult because brain methods do not transfer easily to spinal-cord morphology and intensity-based methods face bias, confounding, and partial-volume effects.
  • CNN rationale: CNNs offer hierarchical representation learning for medical image segmentation rather than relying on manually engineered features.
  • Study objective: The proposed fully automatic framework targets spinal cords and intramedullary MS lesions across varied MRI contrasts, resolutions, pathologies, artifacts, dimensions, and orientations.
  • Data variability: Figure 1 illustrates variability across sites, clinical conditions, image resolution, field of view, contrasts, and lesion appearance.

2. Materials and Methods

The study used heterogeneous multi-site MRI data and a two-stage CNN framework to detect the spinal cord centerline, then segment the cord and/or lesions. Training, inference, evaluation, and implementation were organized across multiple contrasts and clinical cohorts.

  • Data: 1,042 subjects from 30 centers contributed heterogeneous retrospective MRI data spanning multiple clinical conditions.The cohort included healthy controls, MS or suspected MS patients, and other spinal pathologies.
  • Segmentation framework: The framework first detects the spinal cord centerline, then segments the spinal cord and/or lesions along it.The workflow comprises centerline detection followed by segmentation in a volume of interest surrounding the inferred centerline.
  • Spinal cord centerline detection: CNN1 used 2D dilated convolutions to detect the centerline from cross-sectional slices.The architecture adapted U-net by reducing downsampling and replacing conventional convolutions with dilated convolutions.
  • Training and inference: The method was trained separately for each MRI contrast using Dice loss, data augmentation, and independently trained spinal cord and lesion models.Inference thresholded CNN predictions at 0.5 before reconstructing 3D volumes, without additional post-processing.
  • Spinal cord centerline detection: CNN1 predictions were converted into a distance map, from which OptiC inferred and regularised the spinal cord centerline.The centerline then guided extraction of 3D patches for the second-stage segmentation networks.
  • Spinal cord and lesion segmentation: CNN2-SC and CNN2-lesion used 3D CNNs to segment spinal cord and intramedullary lesions in patches surrounding the centerline.The networks drew from 3D U-net designs with reduced depth to limit parameters and memory requirements.

3. Results

The two-stage CNN framework produced strong spinal cord segmentations across heterogeneous clinical and imaging conditions, while MS lesion performance was close to inter-rater agreement and substantially reduced annotation time.

  • Spinal cord segmentation: The method produced encouraging qualitative segmentations across T1-w, T2-w, and T2*-w images from varied sites and clinical conditions.Examples included compressed or atrophied cords, poor contrast, and differing superior-to-inferior coverage.
  • Spinal cord segmentation: 94.6% median Dice versus 87.9% for PropSeg demonstrated superior spinal cord segmentation.The difference was statistically significant (p<0.001).
  • Spinal cord segmentation: 92.9% versus 82.0% Dice in severe cord atrophy showed an advantage over PropSeg in damaged cords.Relative volume difference was -3.6% versus +13.3%.
  • Spinal cord segmentation: 95.2% versus 94.1% median Dice across controls and MS subjects indicated robustness to MS-related pathology.Performance on two unseen sites reached a median Dice of 93.3%.
  • MS lesion segmentation: Lesion discrepancies concentrated near cerebrospinal fluid and grey matter, where partial-volume effects challenged delineation.The network also correctly detected some small lesions and lesions in atrophied cords.
  • MS lesion segmentation: 60.0% median Dice was achieved for automatic MS lesion segmentation across pooled T2-w and T2*-w scans.The median relative volume difference was -14.5%, with lesion-wise precision of 76.9% and sensitivity of 83.3%.
  • MS lesion segmentation: 60.7% versus 56.8% median Dice against majority-voting masks was obtained by raters versus the automatic method.Rater agreement varied substantially, with subject-level differences ranging from 21.0% to 85.0%.
  • MS lesion segmentation: 3.6 minutes versus 18.7 minutes was the average time for automatic versus manual lesion segmentation per subject.The comparison used two volumes per patient on the specified iMac system.

4. Discussion

The framework combines centerline detection with CNN-based segmentation and generalizes across heterogeneous, multi-site spinal cord MRI data. Discussion findings emphasize clinical utility, lesion-segmentation efficiency, and limitations related to training-data requirements, acquisition variability, and inference flexibility.

  • Method: A two-stage CNN first detects the spinal cord centerline, then restricts the second network to a neighborhood for cord or lesion segmentation.The sequential design addresses class imbalance and heterogeneous image features, but makes segmentation dependent on detection quality.
  • Method: A median centerline detection error of 1 mm was reliable enough to cascade detection into segmentation, including scans with low cerebrospinal-fluid/cord contrast.Curve regularization was especially important for maintaining centerline consistency in patients with spinal cord atrophy.
  • Spinal cord segmentation: The method improved spinal cord segmentation over PropSeg, particularly in severe cord compression, where low cerebrospinal-fluid/cord contrast can cause PropSeg leakage.The CNN benefits from a larger spatial view that can use contextual structures such as vertebral edges.
  • Generalization: Performance remained similar on data from new sites, likely reflecting a large, heterogeneous training set spanning diverse scanners, resolutions, orientations, and acquisition parameters.The framework was also consistent across three MRI contrast datasets.
  • Limitations and perspectives: The framework requires extensive manual labeling, may be less reproducible when data cannot be shared, and can perform worse than models trained on homogeneous datasets.Public release and fine-tuning with smaller datasets were proposed to mitigate these constraints.
  • MS lesion segmentation: Automatic lesion segmentation produced results similar to inter-rater results while improving efficiency and reproducibility for large lesion-burden analyses.The reported median Dice was 60.0%, a metric strongly affected by lesion load and lesion size; object-based metrics were also computed.

5. Conclusion

The study presents an original automated method for spinal cord and MS lesion segmentation using two convolutional neural networks. Spinal cord segmentation outperformed a state-of-the-art method on heterogeneous multi-site data, while lesion results were generally within the range of manual segmentations.

  • The proposed method uses a sequence of two convolutional neural networks for spinal cord and MS lesion segmentation.
  • Spinal cord segmentation outperformed a state-of-the-art method on a multi-site, highly heterogeneous clinical dataset.
  • Lesion segmentation results were generally within the range of manual segmentations, although the false positive rate warrants further investigation.
  • The automatic methods are open-source and available in the Spinal Cord Toolbox from version 3.2.2 onward.

Appendix

The appendix summarizes MRI systems, acquisition parameters, and vertebral coverage across contributing sites. The listed material also identifies sites, institutions, study contributions, and relevant disclosures.

  • Table A1 summarizes MRI systems, acquisition parameters, and vertebral coverage across sites.
Loading 1805.06349v2…