Self-supervised Learning for Anomaly Detection in Fundus Image

Since medical data with different characteristics can be observed even with the same disease in a clinical environment, an anomaly detection algorithm should be well applied to medical data that are not seen. Focusing on a fact that an object photograph consists of reflectance and illumination information, we propose a new data augmentation method that can change illumination information for creating a new fundus image by preserving the reflectance information including the disease lesion information. Then our framework which is trained with only normal data during training employs a reconstruction manner with a self-supervised learning technique capable of identifying anomalous images. Based on the reconstruction manner, our model is trained to reconstruct the reflectance image, not the original image to leverage the useful information which is the main component of the fundus image. Furthermore, in order to boost the anomaly detection capability of our proposal, we propose a pretext task for a self-supervised learning manner to reduce intra-class variance by considering the distance of each feature representation. An anomaly score, as a measure to classify the anomalous data, is constructed based on the reconstruction error between the original image and the reconstructed image. In addition, We extensively evaluate our framework on the diabetic retinopathy fundus dataset. The results demonstrate our framework's superiority over the latest state-of-the-art methods.