Towards contrast-agnostic soft segmentation of the spinal cord

TL;DR

This work tackles the problem of contrast-induced variability in spinal cord CSA by proposing a contrast-agnostic soft segmentation framework. It introduces a novel soft-ground-truth generation pipeline across six MRI contrasts and trains a 3D U-Net with aggressive augmentation and an adaptive wing loss to produce calibrated soft segmentations that preserve partial-volume information. The approach reduces CSA variability across contrasts, generalizes to unseen contrasts and pathologies, and demonstrates favorable inference time on CPU compared with state-of-the-art baselines. The method is open-source and integrated into Spinal Cord Toolbox, offering a practical path toward more reliable multi-center CSA biomarkers.

Abstract

Spinal cord segmentation is clinically relevant and is notably used to compute spinal cord cross-sectional area (CSA) for the diagnosis and monitoring of cord compression or neurodegenerative diseases such as multiple sclerosis. While several semi and automatic methods exist, one key limitation remains: the segmentation depends on the MRI contrast, resulting in different CSA across contrasts. This is partly due to the varying appearance of the boundary between the spinal cord and the cerebrospinal fluid that depends on the sequence and acquisition parameters. This contrast-sensitive CSA adds variability in multi-center studies where protocols can vary, reducing the sensitivity to detect subtle atrophies. Moreover, existing methods enhance the CSA variability by training one model per contrast, while also producing binary masks that do not account for partial volume effects. In this work, we present a deep learning-based method that produces soft segmentations of the spinal cord. Using the Spine Generic Public Database of healthy participants ($\text{n}=267$; $\text{contrasts}=6$), we first generated participant-wise soft ground truth (GT) by averaging the binary segmentations across all 6 contrasts. These soft GT, along with aggressive data augmentation and a regression-based loss function, were used to train a U-Net model for spinal cord segmentation. We evaluated our model against state-of-the-art methods and performed ablation studies involving different loss functions and domain generalization methods. Our results show that using the soft segmentations along with a regression loss function reduces CSA variability ($p < 0.05$, Wilcoxon signed-rank test). The proposed spinal cord segmentation model generalizes better than the state-of-the-art methods amongst unseen datasets, vendors, contrasts, and pathologies (compression, lesions), while accounting for partial volume effects.

Figures (28)

• Figure 1: Preprocessing pipeline for soft average segmentations ground truth. (1) Automatic hard spinal cord segmentation using sct_deepseg_sc & manual corrections; (2) Registration to T2w space; (3) Applying each contrast's warping field to bring the segmentation masks to the T2w space; (4) Weighted averaging of segmentations according to each contrast FOV (represented by white rectangles) to create a unique soft GT mask (5) Applying inverse warping fields to bring the unique soft GT to the native space of each contrast.
• Figure 2: Architecture of the proposed SoftSeg model.
• Figure 3: Absolute CSA error between the predictions and GT across each contrast for the proposed model. Scatter plots within each violin represent the individual CSA errors for all participants in the test set. White triangle marker shows the mean CSA error across participants.
• Figure 4: Effect of GT segmentation type (soft vs. hard) on CSA across contrasts. White triangle marker shows the mean CSA across participants.
• Figure 5: Standard deviation of CSA averaged across C2-C3 vertebral levels compared to the baselines (the lower the better). hard_all_SoftSeg refers to the single model trained using all contrasts with hard GT and the SoftSeg training approach Gros2021-ms, hard_all_diceCE_loss refers to the single model trained with the DiceCE loss and hard individual GT, soft_all_diceCE_loss refers to the single model trained with the Dice CE loss and soft GT, soft_per_contrast refers to the mean of 6 individual models trained on 6 contrasts with soft GT, and soft_all refers to the single model trained using all contrasts with soft GT. White triangle marker shows the mean. * $p <0.05$, ** $p < 0.01$, *** $p < 0.001$ (two-sided Bonferroni-corrected non-parametric Wilcoxon signed-rank test).