Analysis of Retinal Vessel Segmentation with Deep Learning and its Effect on Diabetic Retinopathy Classification

Abstract

The success of deep learning methodologies draws a huge attention to their applications in medical image analysis. One of the applications of deep learning is in segmentation of retinal vessel and severity classification of diabetic retinopathy (DR) from retinal funduscopic image. This paper studies U-Net model performance in segmenting retinal vessel with different settings of dropout and batch normalization and use it to investigate the effect of retina vessel in DR classification. Pre-trained Inception V1 network was used to classify the DR severity. Two sets of retinal images, with and without the presence of vessel, were created from MESSIDOR dataset. The vessel extraction process was done using the best trained U-Net on DRIVE dataset. Final analysis showed that retinal vessel is a good feature in classifying both severe and early cases of DR stage.