Multi-labelled Ocular Disease Diagnosis Enforcing Transfer Learning

The leading causes of vision impairment in the working age population today are primarily diseases such as glaucoma, diabetes, etc. Health camps and public health agencies working towards improving eye health among masses engage in activities which require diagnosis on a large scale. This project is aimed at assisting such agencies in effective early diagnosis of these eye diseases by utilizing multi-label CNN-based rapid automated systems to analyze coloured fundus images, thereby mitigating the tedious manual effort associated with clinical diagnosis. Coloured fundus photographs of patients were screened, subjected to various pre-processing techniques - Concatenation, Contrast Limited Adaptive Histogram Equalization and Augmentation; and further classified into 7 labels - Normal, Diabetes, Glaucoma, Cataract, Age related Macular Degeneration, Hypertensive Retinopathy and Pathological Myopia by applying transfer learning using highly effective networks such as VGG-16, InceptionV3 and ResNet50. Performance of each model was evaluated against the Hamming Loss metric. Observed results suggest a significant role of these systems in clinical diagnosis.

I. Introduction

According to the World Health Organization (WHO), more than a billion of the global population today have an entirely preventable near or distance vision impairment that remains untreated [1]. The National Programme for Control of Blindness has emphasized the need for cataract surgical services and refraction services to be augmented, both in quantity and quality, to achieve the goal of eliminating avoidable blindness. The ophthalmologist to population ratio in urban parts of India is 1:25000, declining further to 1:219000 in rural areas [2]. The figures reveal an acute insufficiency of medical personnel in the country.