Automated detection of Hypertensive Retinopathy using few-shot learning

Hypertensive Retinopathy (HR) is a retinal manifestation caused due to persistently raised blood pressure. Computer-aided diagnosis (CAD) plays an important role in the early identification of HR with high diagnostic accuracy, which is time-efficient and demands fewer resources. At present, there are very few computerized systems available for HR detection. Nonetheless, because of the limited number of datasets, there is still room for significant advancement in HR detection. Recently, deep learning has drawn a lot of interest, mainly due to its efficiency has been significantly enhanced. In this work, we develop a novel approach for HR detection based on few-shot learning using a pretrained initial baseline model in which transferable knowledge is obtained for feature embedding on few-shot prediction (limited number of images). It is used to avoid overfitting and to improve generalization on smaller datasets. The proposed baseline model consists of a CNN and LSTM-based HR detection model that can recognize base categories and dynamically generate classification weight vectors for few-shot datasets. The pretrained baseline classifier maximizes the reuse of feature embedding on few-shot datasets, which is comparatively more suitable for smaller datasets than other deep-learning models. In addition, the similarity-based cosine distance classifier followed by the softmax function is used for a few-shot dataset classification. Our experimental findings indicate the effectiveness of the proposed method in HR detection, evaluated on publicly available datasets (including recently released datasets). Therefore, the proposed system can effectively detect HR and can be used by clinicians for referral as well as to facilitate mass screening.