Learning Cataract Severity Using the Contrast Sensetivity Scale: A Thick Data Approach

Cataract image severity classification according to contrast sensitivity chart like VTSC represents a challenge to classical machine learning. Typically cataract images hide significant retina features because of the fogy nature of these images. Contrast sensitivity measurement provides important information on the functioning of the visual system that cannot simple detected by the visual acuity tests. In this paper, we are introducing a framework for segmenting fundus images into regions of salient contrast. The framework start by converting the RGB color channels into joint entropy where it can be used to categorized anchor images into contrast based classes. It has been found that classes like normal fundus, fundus with cataract or fundus having cataract with additional complications like availability of cotton-wool spots or fibrosis. We found that contrast sensitivity scoring correlate nicely with the joint entropy scoring. For this purpose, our last level of the framework employs a triplet loss Siamese neural network that has been trained on few contrast classified fundus images to detect the severity of the cataract. We used a similar contrast sensitivity scale like those used by the optometrist with classification rate that reaches 72%.