A Transformer-Based Knowledge Distillation Network for Cortical Cataract Grading.

Area of Science:

• Ophthalmology
• Computer Vision
• Artificial Intelligence

Background:

• Cortical cataract diagnosis is challenging due to complex lesion features, impacting automated grading accuracy.
• Existing edge detection and deep learning methods show performance degradation with complex cortical opacities and uncertain data.

Purpose of the Study:

• To develop an advanced automated grading system for cortical cataracts.
• To address the limitations of current methods in handling complex opacities and data uncertainty.

Main Methods:

• Proposed a Transformer-based Knowledge Distillation Network (TKD-Net) incorporating zone decomposition for refined feature extraction.
• Introduced sub-scores (location, area, density) and a multi-modal mix-attention Transformer for comprehensive feature learning.
• Implemented Transformer-based knowledge distillation to manage modality missing and uncertain data using a teacher-student model approach.

Main Results:

• TKD-Net demonstrated superior performance compared to state-of-the-art methods in cortical cataract grading.
• Experiments validated the effectiveness of the proposed zone decomposition, sub-scores, and knowledge distillation components.
• The system achieved improved accuracy on slit-lamp images annotated using the LOCS III grading system.

Conclusions:

• TKD-Net offers a robust solution for automated cortical cataract grading, outperforming existing techniques.
• The study highlights the potential of Transformer-based models and knowledge distillation for complex medical image analysis.
• The developed approach effectively handles challenges posed by complex opacities and data variability in clinical settings.