Pathological Myopia Diagnosis based on ResNet SR-init Structured Pruning and Self-learning Distillation

Abstract

This paper proposes a ResNet-based SR-init structured pruning and self-learning distillation system for the diagnosis of pathological myopia. By comparing the performance of VGGNet, GoogLeNet, and ResNet-50 on the iChallenge-PM fundus image dataset, the ResNet-50 model with the best performance was selected based on key indicators such as accuracy and loss value. The SR-init structured pruning technique is then employed to prune filters with minimal performance impact, thereby reducing the model's complexity and computational requirements without significantly affecting its performance, and finally, self-distillation technology is used to further optimize the model. In this process, the model uses its own predicted outputs on the training data as new training labels. These outputs serve as soft labels to retrain the model using the cross-entropy loss function, which aids the model in deeply understanding the complex characteristics of pathological myopia. To prevent overfitting and ensure optimal performance on unseen data, an early stopping strategy was introduced. The final model results performed well, significantly improving the automatic diagnosis efficiency of pathological myopia and reducing the misdiagnosis rate. This study is applicable to resource-constrained medical environments and is of great significance for promoting intelligence and automation in the medical industry.