Unsupervised Retinal Lesion Detection by Learning to Restore Corrupted Fundus Images

Anomaly detection is an important topic in medical image analysis, and some unsupervised deep learning methods have been proposed for lesions detection in medical images. However, these unsupervised methods show poor performance on pixel-level lesion detection in fundus images, since retinal images contain many small lesions and retinal lesions usually show diversity in shapes, sizes, textures, and positions. This paper proposes a new unsupervised framework to solve this problem, in which a high-accuracy anomaly-free image will be reconstructed for a given retinal image, and thus diverse lesions are well separated by comparing these two images. In our framework, an autoencoder augmented by a memory module is forced to reconstruct high-accuracy anomaly-free retinal images from corrupted retinal images with synthetic lesions generated from Perlin noise during training, by which real lesions in anomalous images can be removed on the inference stage. Meanwhile, a false positive suppression algorithm (FPSA) is also proposed to reduce false positives caused by regions with low reconstruction quality. Comparative experiments show that our proposed framework achieves advanced performance and outperforms the state-of-the-art methods on the task of retinal lesion detection from fundus images.