Pan-Ret: a semi-supervised framework for scalable detection of pan-retinal diseases

It has been shown in recent years that a range of optical diseases have early manifestation in retinal fundus images. It is becoming increasingly important to separate the regions of interest (RoI) upfront in the automated classification pipeline in order to ensure the alignment of the disease diagnosis with clinically relevant visual features. In this work, we introduce Pan-Ret, a semi-supervised framework which starts with locating the abnormalities in the biomedically relevant RoIs of a retinal image in an "annotation-agnostic" fashion. It does so by leveraging an anomaly detection setup using parallel autoencoders that are trained only on healthy population initially. Then, the anomalous images are separated based on the RoIs using a fully interpretable classifier like support vector machine (SVM). Experimental results show that the proposed approach yields an overall F1-score of 0.95 and 0.96 in detecting abnormalities on two different public datasets covering a diverse range of retinal diseases including diabetic retinopathy, hypertensive retinopathy, glaucoma, age-related macular degeneration, and several more in a staged manner. Thus, the work presents a milestone towards a pan-retinal disease diagnostic pipeline that can not only cater to the current set of disease classes, but has the capacity of adding further classes down the line. This is due to an anomaly detection style one-class learning setup of the deep autoencoder piece of the proposed pipeline, thus improving the generalizability of this approach compared to usual fully supervised competitors. This is also expected to increase the practical translational potential of Pan-Ret in a real-life scalable clinical setting.