MODALITY-AGNOSTIC LEARNING FOR MEDICAL IMAGE SEGMENTATION USING MULTI-MODALITY SELF-DISTILLATION.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Clinical translation of medical image segmentation models is hindered by the limited availability of all required imaging modalities for a given patient.
• Existing modality-agnostic (MAG) learning trains a single model on all available modalities, remaining input-agnostic for flexible application.

Purpose of the Study:

• To propose a novel framework, MAG learning through Multi-modality Self-distillation (MAG-MS), for robust medical image segmentation.
• To enhance representation learning for individual modalities by distilling knowledge from multi-modality fusion.

Main Methods:

• Developed MAG-MS, a framework that distills knowledge from fused multiple modalities to improve individual modality representation.
• Implemented a self-distillation mechanism within the modality-agnostic learning paradigm.
• Validated the framework on benchmark medical image segmentation datasets.

Main Results:

• MAG-MS demonstrated superior segmentation accuracy compared to state-of-the-art methods.
• The proposed method showed enhanced modality-agnostic (MAG) robustness, performing well with varying modality combinations.
• Experiments confirmed the efficiency of MAG-MS, particularly in scenarios with limited available modalities during testing.

Conclusions:

• MAG-MS offers an adaptable and efficient solution for medical image segmentation, effectively addressing the challenge of limited modality availability.
• The self-distillation approach within a modality-agnostic framework significantly improves segmentation performance and robustness.
• This work advances the clinical translation of deep learning models for medical image analysis by enabling flexible use across different imaging data combinations.