Semi-supervised Liver Segmentation and Patch-based Fibrosis Staging with Registration-aided Multi-parametric MRI

Liver fibrosis poses a substantial challenge in clinical practice, emphasizing the necessity for precise liver segmentation and accurate disease staging. Based on the CARE Liver 2025 Track 4 Challenge, this study introduces a multi-task deep learning framework developed for liver segmentation (LiSeg) and liver fibrosis staging (LiFS) using multiparametric MRI. The LiSeg phase addresses the challenge of limited annotated images and the complexities of multi-parametric MRI data by employing a semi-supervised learning model that integrates image segmentation and registration. By leveraging both labeled and unlabeled data, the model overcomes the difficulties introduced by domain shifts and variations across modalities. In the LiFS phase, we employed a patchbased method which allows the visualization of liver fibrosis stages based on the classification outputs. Our approach effectively handles multimodality imaging data, limited labels, and domain shifts. The proposed method has been tested by the challenge organizer on an independent test set that includes in-distribution (ID) and out-of-distribution (OOD) cases using three-channel MRIs (T1, T2, DWI) and seven-channel MRIs (T1, T2, DWI, GED1-GED4). The code is freely available. Github link: https://github.com/mileywang3061/Care-Liver

💡 Research Summary

This paper presents a dual‑task deep learning framework designed for the CARE Liver 2025 Track 4 challenge, which requires accurate liver segmentation and fibrosis staging from multi‑vendor, multi‑phase MRI data. The authors address two major difficulties: (1) a very limited set of manually annotated liver masks (only the hepatobiliary phase GED4) and (2) substantial domain shifts across modalities (T1, T2, DWI, and four contrast‑enhanced phases GED1‑GED4).
For liver segmentation (LiSeg), they adopt the Better Registration Better Segmentation (BRBS) architecture and extend it to a semi‑supervised setting. The key innovations are: a Weighted Consistency Constraint (WCC) that forces the warped atlas mask to agree with the segmentation prediction, and a Space‑Style Sampling Program (S3P) that generates synthetic image‑mask pairs by interpolating spatial deformations and style (contrast) changes between an atlas and an unlabeled image. These mechanisms enrich the training set with realistic variations and improve pseudo‑label reliability. To cope with intensity heterogeneity across vendors, the original Normalized Cross‑Correlation loss is replaced by a patch‑wise Mutual Information (MI) loss, which is invariant to intensity scaling and better suited for multi‑modal registration. The resulting BRBS‑MI model jointly learns deformable registration and segmentation, propagating the few available GED4 masks to all other sequences. Experiments on the validation set show consistent Dice improvements (up to +10 % points) and reduced Hausdorff distances, especially for T2 and DWI where modality differences are strongest.
For fibrosis staging (LiFS), the authors propose a patch‑based classification pipeline. After obtaining aligned multi‑parametric MRIs and liver masks, overlapping 16 × 16 patches are extracted from the liver region. A simplified ResNet‑18 backbone followed by a multilayer perceptron classifies each patch as “Stage 1‑like” (healthy) or “Stage 4‑like” (severe). Only Stage 1 and Stage 4 subjects are used for training; Stage 2 and Stage 3 cases are reserved for validation and testing, encouraging the network to focus on the most discriminative extremes. Multi‑channel inputs consist of three non‑contrast channels (T1, T2, DWI) or seven contrast‑enhanced channels (adding GED1‑GED4); missing modalities are zero‑filled. During inference, the proportion of Stage 4 patches over all patches is computed. Two thresholds (τ₁ for substantial fibrosis detection, τ₂ for cirrhosis detection) are derived via cross‑validation and applied to map the proportion to probabilities ˆy₁ and ˆy₄, which are then reported as the final stage predictions.
The framework is evaluated on a test set comprising in‑distribution (ID) data from vendors A, B1, B2 and out‑of‑distribution (OOD) data from vendor C. For segmentation, the BRBS‑MI model achieves Dice scores of 93.18 % (ID T1) and 95.03 % (OOD T1), ranking second overall. It attains the best OOD performance on DWI (Dice = 90.41 %) and solid results on other sequences, demonstrating robustness to domain shift. T2 segmentation is comparatively weaker (Dice ≈ 62–76 %), likely due to higher intensity heterogeneity and limited training examples.
For fibrosis classification, the patch‑based approach yields competitive accuracies. In ID cases, it obtains the highest accuracy for Stage 4 vs. 1‑3 (ACC = 75.83 %) and third‑best for Stage 1 vs. 2‑4 (ACC = 77.50 %). In OOD cases, it ranks second for Stage 4 detection (ACC = 58.57 %) and first for Stage 1 detection (ACC = 92.86 %). Corresponding AUC values are also among the top performers, confirming the method’s ability to generalize across unseen vendor data.
Overall, the paper contributes a practical solution to three intertwined challenges: limited annotation, multi‑modal intensity variability, and domain shift. By integrating registration‑aware semi‑supervised segmentation with a simple yet effective patch‑level classifier, the authors deliver a system that is both accurate and interpretable. The code is publicly released, facilitating reproducibility and future extensions to other liver diseases or imaging modalities.