Use of Convolutional Neural Network for Fully Automated Segmentation of Hard Exudates in Retinal Images

Abstract

Diabetic retinopathy (DR), is a complication with diabetes caused by damaged blood vessels in the back of the retina. DR affects 126.6 million people around the world and is the leading cause of blindness. Hard exudates are a type of lesion caused by the damaged blood vessels and are an early marker for DR. In this research, a fully automatic deep learning method has been developed that is able to delineate hard exudate lesions in retinal images. This allows the lesion volume to be calculated and thus determine DR severity. This technology would remove the need for doctors in the diagnosis process, therefore making the diagnosis faster and more accessible to people around the world. Our dataset consisted of 58 images and was used to train a fully convolutional neural network with a U-net architecture. The U-net consists of a contracting path followed by a symmetric expansive path that was used to learn features of the images. These features were then used to differentiate hard exudates from regular tissue allowing them to be segmented. After creating the model 26 images were used for testing. Results of the U-net model showed a Dice similarity coefficient of 67.23 ± 13.60%, a specificity of 99.74 ± 0.25%, and precision of 75.87 ± 18.14% when comparing the algorithm generated images to the manually segmented ground truths. These results show that the model is precisely delineating the hard exudates and therefore is a viable way to diagnose the severity of DR.