Abstract:

Objective To address the challenges of accurately capturing critical features in small?sample diabetic retinopathy （DR） recognition models in real?world applications， and the overly smooth distribution of true and false feature coefficients， we propose an enhanced small?sample DR recognition method based on an improved capsule network. Methods Firstly， the method enhances image feature representation by removing redundant boundary information and employing discrete wavelet transform based on the Haar wavelet function， thereby highlighting critical pathological features. Secondly， the convolutional layer of the capsule network is optimized through a multi?branch architecture to extract multi?scale features from retinal images， while incorporating a convolutional block attention module that is subsequently fed into the capsule layer. Finally， the sigmoid function replaces the softmax function in dynamic routing， thereby improving the model's robustness. Result The enhanced neural network model achieved an accuracy of 98.62% on the Kaggle public dataset following a rigorous selection and preprocessing procedure. Conclusion The enhanced capsule network demonstrated superior precision in identifying diabetic retinopathy within small sample sizes compared to other state?of?the?art algorithms currently available.

Key words: diabetic retinopathy recognition, multi-scale, small sample, capsule network, CBAM