Drashtikon: Extra Ocular Disease Classification

This work outlines the development of a classifier for extra ocular conditions that uses natural, noisy images of faces taken with point-and-shoot cameras. The entire dataset consists of about 7,244 labeled images of patients from Drashti Netralaya Eye Hospital in Gujarat, India. Five diseases were selected for classification: Goonderson Flap, Dermoid Cyst, Strabismus, Ptosis, and Ocular Surface Disease. Histogram of Oriented Gradient feature descriptors were utilized with Support Vector Machines and Logistic Regression. Modern Neural Network architectures were also applied. Bottleneck CNN and Logistic Regression (Balanced) both performed well according to different error measurements. The Bottleneck CNN achieved the highest test accuracy of 77% but Logistic Regression had the highest true positive rate averaged across all classes. Thus, it is possible to develop a classifier from everyday images for a variety of eye diseases. More research is necessary to develop robust predictive capabilities across all diseases.