DRDr II: Detecting the Severity Level of Diabetic Retinopathy Using Mask RCNN and Transfer Learning

TL;DR

This paper presents DRDr II, a hybrid deep learning model that leverages Mask R-CNN and transfer learning to accurately classify diabetic retinopathy severity levels using a large dataset of fundus images.

Contribution

It introduces a novel hybrid approach combining segmentation and transfer learning for DR severity detection, achieving over 92% accuracy.

Findings

Achieved over 92% accuracy in DR severity classification.

Utilized a large, diverse dataset of 35,000+ images.

Enhanced detection by combining segmentation masks with transfer learning.

Abstract

DRDr II is a hybrid of machine learning and deep learning worlds. It builds on the successes of its antecedent, namely, DRDr, that was trained to detect, locate, and create segmentation masks for two types of lesions (exudates and microaneurysms) that can be found in the eyes of the Diabetic Retinopathy (DR) patients; and uses the entire model as a solid feature extractor in the core of its pipeline to detect the severity level of the DR cases. We employ a big dataset with over 35 thousand fundus images collected from around the globe and after 2 phases of preprocessing alongside feature extraction, we succeed in predicting the correct severity levels with over 92% accuracy.