Semantic token-guided hierarchical adversarial knowledge distillation for 3D abdominal organ segmentation.

Deploying computationally intensive 3D models for accurate abdominal organ segmentation remains a significant challenge, particularly in resource-constrained clinical settings. Knowledge distillation (KD) offers a viable path by transferring knowledge from a cumbersome teacher to a compact student network. However, prevailing KD methods are hampered by two key limitations: inadequate modeling of dynamic semantic relationships across multi-scale features, leading to misalignment in low-contrast regions, and an inability to bridge the architectural heterogeneity gap (e.g., Transformer teacher to CNN student), resulting in feature distribution discrepancies. To overcome these issues, we propose STM_HAC-KD, a novel KD framework that synergistically integrates a Semantic Token-guided Multi-scale KD (STM-KD) module and a Hierarchical Multi-scale Patch-consistent Adversarial Alignment KD (HMPA2-KD) module. STM-KD employs learnable, category-aware semantic tokens to establish dynamic cross-scale interactions, effectively correlating shallow structural details with deep semantic context. Complementarily, HMPA2-KD leverages our proposed lightweight 3D Gap Elimination PatchGAN discriminators to adversarially align student feature distributions with the teacher's across multiple scales, thereby eliminating segmentation errors near the boundaries. Comprehensive experiments on the WORD and BTCV datasets demonstrate that STM_HAC-KD consistently outperforms advanced comparative methods, achieving superior Dice Similarity Coefficient (DSC) and significant reductions in 95% Hausdorff Distance (HD95), particularly in boundary-ambiguous regions. This work establishes an efficient and precise paradigm for 3D abdominal organ segmentation, with direct relevance to intelligent clinical decision systems. Our code is available at: https://github.com/oneplus1x/STM_HAC-KD.