Adaptive Knowledge Distillation for Anatomical Segmentation in Pelvic CT Imaging of Prostate Cancer

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Accurate segmentation of prostate and pelvic structures in CT imaging is crucial for prostate cancer treatment planning and staging.
• Challenges in segmentation include low soft-tissue contrast and complex anatomical boundaries.

Purpose of the Study:

• To investigate three knowledge distillation paradigms (voxel-level, region-level, and dynamic weighting) to improve segmentation of prostate and parailiac regions in CT scans.
• To develop a novel fusion approach combining voxel and region-level distillation with uncertainty-aware dynamic weighting.

Main Methods:

• Implemented voxel-level distillation using Kullback-Leibler divergence and region-level distillation using contrastive loss.
• Introduced a fusion method with dynamic weighting to adaptively combine distillation losses.
• Utilized a dual-network architecture (VNet and 3D-ResVNet) for synergistic learning.

Main Results:

• All individual distillation methods consistently improved segmentation accuracy over baseline models on in-house and public datasets.
• The proposed dynamic fusion approach demonstrated robust generalization across different datasets and anatomical structures.
• Effectiveness was shown for segmenting the prostate gland, seminal vesicles, and parailiac regions.

Conclusions:

• Complementary supervision at voxel and region levels significantly enhances the delineation of complex pelvic structures in CT imaging for prostate cancer.
• The developed knowledge distillation strategies offer a robust and practical solution for improving medical image segmentation.