Efficient Medical Image Segmentation Based on Knowledge Distillation.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Convolutional neural networks (CNNs) have advanced medical image segmentation.
• Existing CNNs require substantial computational resources and storage, limiting practical application.
• There is a need for efficient segmentation models suitable for real-world clinical settings.

Purpose of the Study:

• To propose an efficient architecture for medical image segmentation using knowledge distillation.
• To enable lightweight networks to achieve high segmentation performance with reduced resource demands.
• To address the computational and storage limitations of current deep learning segmentation methods.

Main Methods:

• Implemented knowledge distillation from a well-trained teacher network to a lightweight student network.
• Devised a novel distillation module specifically for medical image segmentation to transfer semantic region information.
• The module focuses on transferring internal semantic region information, avoiding ambiguous boundary issues.

Main Results:

• The lightweight network achieved significant improvements in segmentation capability, up to 32.6%.
• The distilled network maintained runtime efficiency and portability during inference.
• Performance was validated on public CT datasets (LiTS17 and KiTS19).

Conclusions:

• The proposed knowledge distillation method effectively enhances lightweight medical image segmentation networks.
• This approach offers a valuable solution for scenarios demanding high operating speed and low storage usage.
• The method demonstrates the potential for efficient and accurate medical image analysis in resource-constrained environments.