Volumetric Medical Image Segmentation Through Dual Self-Distillation in U-Shaped Networks.

Area of Science:

• Medical Image Analysis
• Deep Learning for Medical Imaging
• Computer Vision

Background:

• U-shaped networks are effective for medical image segmentation.
• Existing methods may have limitations in capturing complex anatomical details.

Purpose of the Study:

• To introduce a novel Dual Self-Distillation (DSD) framework for U-shaped networks.
• To improve volumetric medical image segmentation performance.

Main Methods:

• Proposed a Dual Self-Distillation (DSD) framework integrated into U-shaped networks.
• DSD distills knowledge from ground-truth labels to decoder layers and between network layers.
• Applied DSD to state-of-the-art U-shaped backbones for 3D medical image segmentation.

Main Results:

• Significant improvements in Dice Similarity Score (average increase of 2.82% for cardiac, 4.53% for brain tumor, 1.3% for hippocampus).
• Reduced Hausdorff distance (average decrease of 7.15 mm for cardiac, 6.48 mm for brain tumor, 0.76 mm for hippocampus).
• Achieved these gains with negligible increase in parameters and training time.

Conclusions:

• The DSD framework is a generalizable training strategy that enhances U-shaped network performance.
• DSD leads to significant quantitative and qualitative improvements in segmenting various 3D medical images.
• The proposed method offers a computationally efficient approach to boost medical image segmentation accuracy.