CDNet: A collaboration and disentanglement network for cross-domain joint grading of Diabetic Retinopathy and Diabetic Macular Edema

Diabetic Retinopathy (DR) and Diabetic Macular Edema (DME) are two common complications related to diabetes and the leading causes of blindness. Timely diagnosis and treatment can decrease the risk of vision loss resulted from DR and DME. Automatic identifying and grading of DR and DME are critical to improving the accuracy and efficiency of early diagnosis of these diseases. Compared with grading these two diseases with separated models, joint grading of DR and DME obtained improved performance. However, current approaches suffer from performance degradation on cross-dataset experiment. To cope with this problem, we propose a collaboration and disentanglement network (CDNet), which implements a novel “collaboration first, and disentanglement later” approach in this paper. CDNet first extracts the basic fundus image features by a unified backbone to characterize the correlation between the abnormalities of DR and DME. Then, a multi-scale attention block is implemented to extract disentangled DR and DME features. To ensure better separation of disentangled features, a mutual information estimator is incorporated to provide quantitative assessment on the disentanglement performance. By minimizing mutual information between the disentangled features, CDNet is capable of adapting to different properties of cross-domain data. This novel learning scheme helps CDNet learn robust and discriminative features for cross-domain DR and DME grading. We conducted intra-dataset and cross-dataset experiments on three widely used benchmarks: Messidor, IDRiD and APTOS. Compared to the state-of-the-art approaches, the joint grading accuracy of CDNet improved from 46% to 56% on cross-dataset experiment, which demonstrated the superiority of CDNet.