Related Experiment Video
Updated: May 8, 2026
04:48
Swin-PSAxialNet: An Efficient Multi-Organ Segmentation Technique
Published on: July 5, 2024
Semantic token-guided hierarchical adversarial knowledge distillation for 3D abdominal organ segmentation
Xiangchun Yu1, Guangjun Zhu1, Ata Jahangir Moshayedi1
1Jiangxi Provincial Key Laboratory of Multidimensional Intelligent Perception and Control, School of Information Engineering, Jiangxi University of Science and Technology, Ganzhou, 341000, China.
Summary
This study introduces STM_HAC-KD, a new knowledge distillation (KD) framework for efficient 3D abdominal organ segmentation. It improves accuracy, especially in challenging low-contrast areas, by better aligning complex models.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Computer Vision
Background:
- Accurate 3D abdominal organ segmentation is crucial but challenging in resource-limited clinical settings.
- Existing knowledge distillation (KD) methods struggle with multi-scale feature relationships and architectural differences between teacher and student models.
- These limitations lead to misalignments in low-contrast regions and feature discrepancies, hindering segmentation performance.
Purpose of the Study:
- To develop a novel KD framework, STM_HAC-KD, that addresses the limitations of current methods for 3D abdominal organ segmentation.
- To improve the accuracy and efficiency of deploying computationally intensive 3D segmentation models in clinical environments.
- To overcome challenges related to dynamic semantic relationships and architectural heterogeneity in knowledge transfer.
Main Methods:
- Proposed STM_HAC-KD framework integrating Semantic Token-guided Multi-scale KD (STM-KD) and Hierarchical Multi-scale Patch-consistent Adversarial Alignment KD (HMPA^2-KD).
- STM-KD utilizes learnable, category-aware semantic tokens for dynamic cross-scale feature interactions.
- HMPA^2-KD employs lightweight 3D Gap Elimination PatchGAN discriminators for adversarial alignment of multi-scale features.
Main Results:
- STM_HAC-KD significantly outperformed advanced comparative methods on the WORD and BTCV datasets.
- Achieved superior Dice Similarity Coefficient (DSC) and substantial reductions in 95% Hausdorff Distance (HD95).
- Demonstrated particular effectiveness in boundary-ambiguous regions and low-contrast areas.
Conclusions:
- STM_HAC-KD establishes an efficient and precise paradigm for 3D abdominal organ segmentation.
- The framework effectively bridges the architectural heterogeneity gap and models dynamic semantic relationships.
- This work has direct relevance for developing intelligent clinical decision systems for medical imaging analysis.