Semi-Supervised Framework for Dual Encoder Attention Network: Classification of Retinopathy in Optical Coherence Tomography Images

Abstract

Due to the high inter-class similarity and intra-class differences in retinal optical coherence tomography (OCT) images, subtle pathological attributes can serve as important dis-criminative clues, and semi-supervised medical image tasks still face challenges. To address this challenge, a new semi-supervised framework for medical image classification is proposed. This method introduces a deep feature fusion classification method based on dual encoders. The different branches of the encoder can extract different feature information, and their features are fused to achieve information complementarity, so that the model can learn subtle pathological attributes in medical images. To improve the model's discrimination and classification performance, an adaptive attention (AA) mechanism is introduced to capture key regions. The proposed method is evaluated on a public OCT dataset. Only 10% of labeled data is used for training, achieving an accuracy of 96.02 %, which is equivalent to the performance of fully supervised models trained with 100% labeled data. Experimental results show that the performance of this method is better than that of supervised benchmark methods and other semi-supervised methods.