End-to-end regression of clinical parameters from retinal images

Abstract

Retinal images are widely used by medical doctors to diagnose eye-related diseases. Unfortunately, the human factor restricts possibilities of full analysis of fundus. Therefore, implementing an automated algorithm for analysing retinal images is beneficial. In recent years, convolutional neural network models for image analysis have been achieving significant results on different datasets. This work focuses on improving training convergence and prediction performance of DenseNet-201, Inception-ResNet-v2 and NASNet-Large deep learning models. Special attention is paid to visual regression for determining the connection between images and clinical parameter data. Based on the experiments, the DenseNet-201 network is stated as the most suitable model for the age and Arteriovenous Ratio prediction task. Visualisation of activations of each network is presented and analysed. For obtaining considerable results in prediction and visualisation tasks more training data and improvements in the approach are needed.