Enhanced Multistructure Segmentation in 3D Whole-body MRI: A Comparative Evaluation of Deep Learning Architectures

Quantitative whole-body MRI relies on accurate delineation of multiple anatomical structures, yet manual labeling is slow and variable. We evaluate AISHANet, a deep learning model for multistructure 3D segmentation, and compare it with SegResNet, UNETR, and UNet using the same dataset split and evaluation protocol. The task covers 14 muscle groups and both lungs.

The dataset includes 100 whole-body DIXON T1 axial volumes acquired on a 3T Philips scanner (one volume per patient) with reference annotations produced by five expert radiologists. We used 80 volumes for training, 12 for validation, and 8 for testing. Performance was assessed with Dice Similarity Coefficient (DSC), directed Hausdorff distance, Sensitivity, ROC AUC, and F1-score. Metrics were computed per patient, macro-averaged across the 16 structures, and summarized as mean ± standard deviation (SD) across test patients.

AISHANet obtained the highest overall scores, with a mean DSC 0.871 ± 0.017, directed Hausdorff distance in millimeters 24.11 ± 10.92, sensitivity 0.894 ± 0.047, ROC AUC 0.947 ± 0.024, and F1-score 0.871 ± 0.061. The best performance was observed in larger muscle groups (gluteus, thighs, calves), where DSC exceeded 0.88.

While AISHANet consistently outperformed the baselines, performance decreased in anatomically challenging regions (abdomen and back), which are affected by lower contrast and thinner structures in axial views. Across models, we observed different failure modes: SegResNet tended to produce smoother masks, UNETR reduced isolated false positives, and UNet showed higher sensitivity to anatomical variability.

Under a controlled, single-protocol comparison, AISHANet provided the highest overall accuracy for multistructure whole-body MRI segmentation in this dataset. Remaining errors in low-contrast and anatomically complex regions motivate future work on improving robustness and validating performance across additional imaging settings, including other MRI scanner manufacturers and additional MRI sequence types.