2sRanking-CNN: A 2-stage ranking-CNN for diagnosis of glaucoma from fundus images using CAM-extracted ROI as an intermediate input

TL;DR

This paper introduces a two-stage ranking-CNN for glaucoma diagnosis from fundus images, utilizing CAM-extracted ROI as an intermediate input, significantly improving accuracy and sensitivity over existing methods.

Contribution

It presents a novel 2-stage ranking-CNN that incorporates CAM-based ROI masks, enhancing classification accuracy and sensitivity for glaucoma detection.

Findings

10% accuracy improvement over existing models

20% sensitivity increase for suspicious cases

ROI overlaps with physician diagnostic criteria

Abstract

Glaucoma is a disease in which the optic nerve is chronically damaged by the elevation of the intra-ocular pressure, resulting in visual field defect. Therefore, it is important to monitor and treat suspected patients before they are confirmed with glaucoma. In this paper, we propose a 2-stage ranking-CNN that classifies fundus images as normal, suspicious, and glaucoma. Furthermore, we propose a method of using the class activation map as a mask filter and combining it with the original fundus image as an intermediate input. Our results have improved the average accuracy by about 10% over the existing 3-class CNN and ranking-CNN, and especially improved the sensitivity of suspicious class by more than 20% over 3-class CNN. In addition, the extracted ROI was also found to overlap with the diagnostic criteria of the physician. The method we propose is expected to be efficiently applied to any medical data where there is a suspicious condition between normal and disease.