DR-HIPI: Performance Evaluation of Retinal Images for DR Lesion Segmentation Using the HIPI Architecture

This chapter proposes a framework that examines possible categories of image modalities available for diabetic retinopathy (DR) and recommends the type of image modality that is better to perform DR segmentation and appropriate post-segmentation machine vision analysis for better DR diagnosis. The proposed system consists of pre-processing step, which uses an order statistics filter to eliminate the noise from the acquired retinal images. The pre-processed images will be sent to the Hadoop Image Processing Interface (HIPI) architecture, implemented using the map-reduce framework. Necessary map and reduce algorithms were implemented to measure the mean pixel intensity values for each type of image modality. For comparison, the proposed system analysis employs 50 color fundus photographs (CFP), 30 fluorescein angiography (FA), and 50 optical coherence tomography (OCT). The average pixel intensity values for CFP are close to 0.31, 0.21, and 0.19 for the RGB channel. The FA values obtained were relative to 0.19, 0.08, and 0.04 for the RGB channel. Similarly, the values for OCT were 0.04, 0.04, and 0.04 for the RGB channel. In comparison to FA and OCT, our proposed work has shown that CFP produces better results. With the proposed map and reduce algorithms, any other type of fundus image could be implemented in the future. The methodology we have implemented using the HIPI architecture is not found in previous research findings in DR, which is the primary motivation for carrying out our work.