Multi-scale contextual learning for medical image segmentation via dual distillation

Area of Science:

• Medical Image Analysis
• Deep Learning for Healthcare
• Computational Pathology

Background:

• Hybrid models combining Convolutional Neural Networks (CNNs) and transformers show promise for medical image segmentation by integrating multi-scale representations.
• However, these hybrid approaches often suffer from high computational and space complexities, limiting their clinical applicability.

Purpose of the Study:

• To develop a computationally efficient hybrid model for medical image segmentation.
• To address the limitations of high complexity in existing CNN-transformer fusion models for resource-constrained clinical settings.

Main Methods:

• A novel model utilizing a dual distillation scheme to leverage complementary CNN and transformer advantages without sacrificing efficiency.
• Introduction of a multi-scale prior-knowledge distillation (MPD) module for effective distillation of multi-scale knowledge from transformer features.
• Development of an efficient and robust Selective Fusion module within the student network to complement the knowledge distillation process.

Main Results:

• Extensive evaluation on SipakMed and ISIC 2017 datasets against fourteen network frameworks.
• The proposed model significantly outperforms existing methods in mean Intersection over Union, mean Dice coefficient, and mean average symmetric surface distance.
• Demonstrated reduction in computational resources, including model parameters and floating-point operations per second.

Conclusions:

• The developed method efficiently integrates CNN and transformer strengths for medical image segmentation.
• Achieves superior segmentation accuracy and reduced computational complexity, confirming its suitability for clinical applications.