OrdPrune-KD: An Ordinal-Consistency-Based Model Compression Framework for Diabetic Retinopathy Grading

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Diabetic retinopathy (DR) grading requires accurate and efficient models.
• Existing model compression techniques may not fully leverage ordinal relationships in DR grading.
• Deploying lightweight yet high-performing DR grading systems is crucial for clinical applications.

Purpose of the Study:

• To propose OrdPrune-KD, an ordinal-consistency-driven model compression framework for diabetic retinopathy grading.
• To integrate grade-aware structured pruning and Earth Mover's Distance (EMD)-based knowledge distillation.
• To incorporate ordinal priors into both model compression and knowledge transfer stages.

Main Methods:

• Developed OrdPrune-KD, a novel framework combining structured pruning and knowledge distillation.
• Incorporated ordinal priors into the model compression and knowledge transfer processes.
• Utilized Earth Mover's Distance (EMD) for knowledge distillation.

Main Results:

• The proposed framework achieved a favorable balance between model compactness and predictive performance.
• A 77% parameter reduction was achieved with the student model showing competitive performance (QWK) and strong high-risk sensitivity.
• Performance gains were attributed to the ordinal-aware design, not output formulation differences.

Conclusions:

• OrdPrune-KD offers an effective and deployable solution for lightweight DR grading systems.
• The ordinal-aware design enhances both compression and predictive performance.
• This framework addresses the need for efficient and accurate DR grading tools.