PaperlayerView exact revision on arXiv

Source-linked AI summary

Spinal cord gray matter segmentation using deep dilated convolutions

Christian S. Perone, Evan Calabrese, Julien Cohen-Adad

arXiv:1710.01269v1cs.CV

TL;DR

Spinal cord gray matter changes are associated with neurological disorders and disability, making automated assessment relevant to spinal cord studies. This paper proposes an end-to-end deep-learning segmentation method for in vivo and ex vivo MRI, reporting strong challenge results and a substantial parameter reduction relative to U-Net.

  • Problem

    Spinal cord gray matter changes are linked to neurological disorders and gray matter atrophy is a relevant disability biomarker in amyotrophic lateral sclerosis, motivating automatic segmentation.

    Source: §1 · p. 1

  • Method

    The paper proposes fully automated end-to-end spinal cord gray matter segmentation using a deep-learning architecture based on Atrous Spatial Pyramid Pooling (ASPP) [7][6].

    Source: §1 · p. 2

  • Results

    The method achieved better results in 8 out of 10 SCGM Challenge metrics and reduced parameters by more than 6 times compared with U-Net.

    Source: §5 · pp. 8, 11

  • Takeaways & Limitations

    The approach is reported to work on both in vivo and ex vivo MRI acquisitions, including ex vivo high-resolution volumes with only a few annotated axial slices.

    Source: §1 · p. 2

    The current approach is limited to 2D slices; adding 3D context is expected to increase memory consumption.

    Source: §5 · p. 11

Abstract

from arXiv · show

Gray matter (GM) tissue changes have been associated with a wide range of neurological disorders and was also recently found relevant as a biomarker for disability in amyotrophic lateral sclerosis. The ability to automatically segment the GM is, therefore, an important task for modern studies of the spinal cord. In this work, we devise a modern, simple and end-to-end fully automated human spinal cord gray matter segmentation method using Deep Learning, that works both on in vivo and ex vivo MRI acquisitions. We evaluate our method against six independently developed methods on a GM segmentation challenge and report state-of-the-art results in 8 out of 10 different evaluation metrics as well as major network parameter reduction when compared to the traditional medical imaging architectures such as U-Nets.

1 Introduction

Spinal cord gray matter changes are clinically relevant, but their segmentation remains difficult amid image and data variability. The paper introduces an end-to-end automated deep-learning approach and evaluates it against independently developed methods.

1 Introduction

  • Spinal cord gray matter changes have been linked to neurological disorders, and gray matter atrophy is a relevant disability biomarker in amyotrophic lateral sclerosis.In multiple sclerosis, spinal cord gray matter was reported as the strongest correlate of disability in multivariate models.Source: §1 · p. 1
  • Automated assessment is important for studying in vivo and ex vivo spinal cords, supporting co-registration and spatial normalization, and reducing the time burden of longitudinal delineation.Source: §1 · p. 1
  • Gray matter segmentation remains challenging because surrounding tissue intensities are inconsistent and images may contain artifacts or pathology-induced contrast changes.The section also identifies differences in acquisition protocols, pixel sizes, gold-standard segmentations, and evaluation metrics as sources of complexity.Source: §1 · p. 1
  • Medical imaging deep learning faces costly acquisition and specialist annotation, variable MRI equipment and protocols, and limited data availability under privacy or ethics constraints.Source: §1 · p. 2
  • The proposed end-to-end deep-learning pipeline targets fully automated spinal cord gray matter segmentation and is evaluated against six independently developed methods.The authors report state-of-the-art results on many metrics in an independent in vivo evaluation and generalization to ex vivo high-resolution MRI with only a few annotated slices.Source: §1 · p. 2

2 Related Work

Prior spinal cord segmentation methods include semi-automated and fully automated approaches, spanning intensity-, surface-, and image-based methods. The section describes representative GM methods and the challenge dataset used for their comparison.

2 Related Work

  • Spinal cord segmentation methods are grouped as semi-automated or fully automated, with intensity-, surface-, and image-based approaches among the categories.Source: §2 · p. 2
  • Semi-supervised VBEM models MRI intensities with a Gaussian mixture and estimates parameters by variational EM, using warped reference anatomy with optional manual labels.Source: §2 · p. 2
  • SCT selects similar preprocessed dictionary slices in PCA space and fuses their manual segmentations; it incorporates vertebral-level information and linear intensity normalization for multisite data.Source: §2 · p. 2
  • JCSCS combines OPAL cord detection with STEPS GM segmentation, propagating registered atlases slice-wise before fusing selected templates.Source: §2 · p. 3
  • MGAC uses an externally estimated cord boundary and a transformed GM template as the initial guess for an active contour.Source: §2 · p. 3
  • GSBME is a semi-automatic supervised method that thresholds slices using maximum GM-plus-WM intensity entropy, then rejects outliers using morphological features.Source: §2 · p. 3
  • Deepseg uses a U-Net-like contracting and expanding network with shortcuts, unpooling and deconvolution, pretraining, and a weighted loss balancing GM and non-GM errors.Source: §2 · p. 3
  • The SCGM Challenge provides a public standard dataset assembled by four spinal-cord imaging groups to compare six independently developed GM segmentation methods.The paper compares its method with the previously described methods on this dataset.Source: §2 · p. 4

3 Methods and Materials

The paper focuses on a simple deep-learning method that is trained end-to-end and is intended to generalize from a small number of labeled axial slices.

3 Methods and Materials

  • Many prior GM segmentation methods rely on registered templates or atlases, arbitrary similarity metrics, or complex pipelines not optimized end-to-end.Source: §3 · p. 4
  • The proposed method is a simple deep-learning approach trained end-to-end and designed to generalize from a small number of labeled 2D axial slices of a 3D MRI volume.Source: §3 · p. 4
  • The approach is framed as a contrast to earlier pipelines that use registrations, external atlases or templates, and label-fusion stages.Source: §3 · p. 4

3.1 Note on U-Nets

U-Net uses contracting and expanding paths with shortcuts to combine semantic aggregation and spatial recovery. The section notes parameter growth and potential gradient-flow issues in its deeper layers.

3.1 Note on U-Nets

  • U-Net uses contracting and expanding paths to aggregate semantic information and recover spatial detail through shortcut connections.Source: §3.1 · p. 4
  • The two-path U-Net architecture expands the parameter space, and the authors describe problematic gradient flow in its high-level layers [41].Source: §3.1 · p. 4
  • Visualized U-Net bottom-layer features were noisy, while low-level features showed meaningful patterns; removing bottom layers produced the same or sometimes better performance.Source: §3.1 · p. 4

3.2 Proposed method

The proposed spinal-cord gray-matter segmenter adapts an ASPP architecture with dilated convolutions, a Dice loss, and extensive data augmentation. Its design removes decimation operations to retain translational equivariance and reduce parameters.

3.2 Proposed method

  • Dilated convolutions expand the receptive field without proportionally increasing parameters, while avoiding pooling helps preserve translational equivariance for dense prediction.The method removes decimation operations and trades depth to reduce memory use and parameters.Source: §3.2 · p. 4
  • The ASPP-based network combines standard and dilated convolutions with parallel branches, then produces a dense prediction map.Its six parallel branches use 1x1 convolutions, dilations at rates 6, 12, 18, and 24, and global average pooling.Source: §3.2 · pp. 4–5
  • Dice Loss is used instead of cross-entropy because gray-matter segmentation is highly imbalanced; the authors report better results than weighted cross-entropy.The epsilon term stabilizes the loss, while the weighted cross-entropy used by [33] is described as harder to optimize because of its weighting hyper-parameter.Source: §3.2 · p. 5
  • Training augmentation includes rotation, shifting, scaling, flipping, noise, and elastic deformation.These strategies address the need for regularization and augmentation in comparatively small medical-image datasets.Source: §3.2 · p. 5

3.2.1 U-Net architecture

  • The U-Net comparison used a 14-layer, 2D convolutional model with the same training protocol and loss, but more aggressive augmentation to address overfitting.The authors selected the best-performing architecture after extensive exploration.Source: §3.2.1 · p. 5

3.3 Data sets

Evaluation uses the SCGM Challenge dataset and an ex vivo high-resolution spinal-cord acquisition. The described pipeline resamples axial slices to a common pixel size, normalizes intensity, and then runs network inference.

3.3.1 Spinal Cord Gray Matter Challenge

  • The SCGM Challenge dataset contains 80 healthy subjects from four centers, split into 40 training and 40 hidden-test subjects, with four independent expert masks per labeled slice.Its three MRI systems and acquisition parameters yield voxel sizes from 0.25x0.25x2.5 mm to 0.5x0.5x5.0 mm.Source: §3.3.1 · p. 5
  • Before inference, axial slices are resampled to a common pixel-size space and intensity-normalized.The pipeline then applies the network to the preprocessed images.Source: Figure 4 · p. 6
  • Training-set misclassified voxels were reported, but evaluations retained the original dataset for fair comparison.This means the reported evaluations did not use a corrected version of those training labels.Source: §3.3.1 · p. 6

3.3.2 Ex vivo high-resolution spinal cord

  • The ex vivo evaluation uses an entire human spinal cord acquired at 7T, comprising 4,676 axial slices at 100 µm isotropic resolution.The acquisition took approximately 120 hours.Source: §3.3.2 · p. 6

3.4 Training Protocol

The training protocols cover the SCGM Challenge and an ex vivo high-resolution data set, with augmentation and a poly learning-rate schedule used in both. The authors also interpret multiple rater masks as contour smoothing and leave explicit contour-label smoothing for future work.

3.4.1 Spinal Cord Gray Matter Challenge

  • The SCGM Challenge volumes were resampled to 0.25×0.25 mm, center-cropped to 200×200 pixels, and mean-centered and standard-deviation normalized.Eight subjects, two per site, were held out for validation; the remaining subjects were used for training, without external data or vertebral information.Source: §3.4.1 · p. 6
  • Training used Adam with initial learning rate η = 0.001, dropout rate 0.4, and the poly learning-rate policy with p = 0.9 for 1000 epochs.Source: §3.4.1 · pp. 6–7

3.4.2 Inter-rater variability as label smoothing regularization

  • Using all four independently created masks as gold standards was interpreted as label smoothing because rater differences usually occur at the GM border.The authors relate this to less-confident contour predictions and propose that artificially smoothing target labels at disagreement-prone contours could potentially improve generalization; they leave this exploration to future work.Source: §3.4.2 · p. 7

3.4.3 Ex vivo high-resolution spinal cord

  • The ex vivo split comprised 15 evenly spaced training slices from 4676 axial slices, plus 7 evenly spaced validation slices and 8 test slices distributed across the volume.Augmentations included rotation, shift, scaling, channel shift, flipping and elastic deformation; random search was not employed to avoid overfitting due to dataset size.Source: §3.4.3 · p. 7
  • The ex vivo model trained in approximately 2 hours and segmented 4676 axial slices in approximately 25 seconds.Training used 600 epochs and the same τ = 0.999 threshold used for the SCGM task.Source: §3.4.3 · p. 7

4 Results

On the SCGM Challenge, the method achieved state-of-the-art results on 8 of 10 metrics and outperformed four of six earlier methods across all metrics. On ex vivo data, it outperformed U-Net on 6 of 8 metrics with sixfold fewer parameters; qualitative comparisons also describe differences in segmentation errors.

4.1.1 Qualitative Evaluation

  • The method captured GM anatomy in four-center SCGM test samples, including under blur, and segmented the gray commissure where Deepseg did not.The comparison describes Deepseg as using a U-Net-like structure with pre-training and 3D-wise training.Source: §4.1.1 · p. 8

4.1.2 Quantitative Evaluation

  • State-of-the-art results on 8 of 10 metrics; the method surpassed 4 of 6 previously developed methods on all metrics.It also performed better than Deepseg on 8 of 10 metrics despite using neither 3D convolutions, pre-training, nor threshold optimization.Source: §4.1.2 · p. 8
  • The authors report state-of-the-art performance across overlap, distance, and statistical metrics, while TNR and PPV were not the best-performing metrics.They hypothesize that sharp Dice-Loss decision boundaries contributed to suboptimal PPV and note that threshold optimization was not performed.Source: §4.1.2 · p. 8

4.2.1 Qualitative Evaluation

  • In ex vivo comparisons, U-Net predictions extended into CSF near the dorsal horn and oversegmented toward white matter; the proposed method was more contained and performed well.Source: §4.2.1 · p. 8

4.2.2 Quantitative Evaluation

  • The proposed method achieved better results on 6 of 8 ex vivo metrics and used 6× fewer parameters than U-Net.Even with dropout 0.4, U-Net overfit during training, leading the authors to use more aggressive augmentation, especially for shifting.Source: §4.2.2 · p. 8

5 Discussion

The method outperforms six independently developed methods on 8 of 10 SCGM Challenge metrics, while using an end-to-end pipeline and substantially fewer parameters than U-Net. The discussion also notes a 2D-slice scope and the memory cost of adding 3D context.

5 Discussion

  • The method achieves better results than six independently developed methods on 8 of 10 SCGM Challenge metrics.Table 2 reports the challenge metrics and marks the best results; the discussion identifies the 8-of-10 comparison.Source: §5 · p. 8, Table 2 · p. 11
  • Figure 5 shows input, majority-vote gold standard, and method output for the same axial slice; Figure 6 presents test-set results across evaluated challenge metrics.Figure 5 samples subject 11 from each site.Source: Figure 5 · p. 9, Figure 6 · p. 9
  • Figure 7 renders ex vivo lumbosacral gray matter in orange, while Figure 8 contrasts U-Net and proposed-method samples across sacral, thoracic, and cervical regions.Green rectangles in Figure 8 mark U-Net oversegmentation.Source: Figure 7 · p. 10, Figure 8 · p. 10
  • Unlike approaches with separate registration, external atlas or template data, and label-fusion stages, the proposed method optimizes its prediction pipeline end to end.The discussion describes the entire pipeline as optimized using backpropagation and gradient descent.Source: §5 · pp. 8, 11
  • Compared with U-Net on ex vivo high-resolution data, the method performs better on many metrics while reducing parameters by more than 6 times.The authors interpret this parameter efficiency alongside performance as a more compact description of data regularities.Source: §5 · p. 11, Table 3 · p. 11
  • The current approach is limited to 2D slices; adding 3D context is proposed as possible, but would increase memory consumption.The authors also suggest semi-supervised learning as a possible extension, citing smoothness in axial slices, especially in ex vivo high-resolution MRI.Source: §5 · p. 11

Related papers

Loading 1710.01269v1…