Weakly-Supervised Lesion-Aware and Consistency Regularization for Retinitis Pigmentosa Detection from Ultra-Widefield Images

摘要

Retinitis pigmentosa (RP) is one of the most common retinal diseases caused by gene defects, which can lead to night blindness or complete blindness. Accurate diagnosis and lesion identification are significant tasks for clinicians to assess fundus images. However, due to some limitations, it is still challenging to design a method that can simultaneously diagnose and accomplish lesion identification so that the accurate lesion identification can promote the accuracy of diagnosis. In this paper, we propose a method based on weakly-supervised lesion-aware and consistency regularization to detect RP and generate lesion attention map (LAM). Specifically, we extend global average pooling to multiple scales, and use multi-scale features to offset the gap between semantic information and spatial information to generate a more refined LAM. At the same time, we regularize LAMs with different affine transforms for the same sample, and force them to produce more accurate predictions and reduce the overconfidence of the network, which can enhance LAM to cover lesions. We use two central datasets to verify the effectiveness of the proposed model. We train the proposed model in one dataset and test it in the other dataset to verify the generalization performance. Experimental results show that our method achieves promising performance.