Task-Adaptive Multi-Prototype Prototypical Networks for keratoconus diagnosis under unbalanced small samples

Objective: This study aims to develop and evaluate a meta-learning-based intelligent diagnostic algorithm for Keratoconus (KC), addressing limited training samples and class imbalance in corneal topography analysis. Early detection of KC remains clinically challenging yet crucial for timely intervention. Methods: This study proposes a Task-Adaptive Multi-Prototype Prototypical Network (TA-MPN) architecture with a Swin Transformer backbone. The approach features: (1) a multi-prototype representation strategy (2-4-2 configuration for Normal, E-KC (incipient stage), and KC (complete stage) classes), (2) a task-level prototype adaptation mechanism, and (3) a balanced task sampling strategy with dual-loss optimization. The dataset comprised 462 corneal topographies with patient-level stratification into training (70%), validation (15%), and test (15%) sets. Results: The proposed model achieved 96.9% overall accuracy on the independent test set, with class-specific accuracies of 97.5% (Normal), 96.7% (E-KC), and 96.7% (KC). This performance represents a 1.1% improvement over state-of-the-art methods, with particularly enhanced performance in detecting E-KC cases. Both sensitivity and specificity exceeded 96% for all three categories. Conclusion: The meta-learning approach demonstrates high diagnostic performance under small-sample and class-imbalanced conditions. The high sensitivity (96.7%) for E-KC detection offers significant clinical value for timely intervention. The proposed framework provides a promising solution for medical image analysis with limited and imbalanced data.