A Semi-supervised Convolutional Autoencoder Framework using Transfer-learning based Pseudo-label Generation for Glaucoma Diagnosis

Glaucoma is a retinal disease that causes irreversible damage to the Optic Nerve Head (ONH) and can result in partial or complete blindness if left untreated. Early detection and treatment of the disease can alleviate such serious effects. Computer Aided Diagnosis (CAD) is gaining importance as an aid to clinicians in scenarios such as mass screening. With the advances in computational resources, the use of Deep Learning (DL) techniques in CAD systems is increasingly preferred due to DL’s capability of extracting high-level features automatically. However, the training of a DL model requires a large number of annotated data. Nevertheless, data pertaining to medical domain is mostly unlabelled, with a relatively small percentage of labelled data, and glaucoma is no exception. The proposed work aims to develop a DL-based CAD system for glaucoma diagnosis using retinal fundus images by integrating approaches including semi-supervised learning, transfer learning and pseudo-label generation for effective use of unlabelled data. Specifically, a Convolutional Autoencoder (CAE) is first trained in an unsupervised manner using unlabelled data. Secondly, a transfer learning model is fine-tuned with limited labelled data available and used for predicting pseudo-labels of the unlabelled cases, which are fed to the autoencoder model for feature extraction. Subsequently, classification is performed for glaucoma detection via a Support Vector Machine (SVM) classifier. The experiments conducted on the combined dataset comprising fundus images from ORIGA and DHRISTI-GS has yielded an accuracy of 90% for the proposed approach by utilizing a large unlabelled dataset and minimal labelled data, showing the efficacy of the approach in the absence of abundant labelled data.