DR-AIVis: A Hybrid Approach for Diabetic Retinopathy Detection Using U-Net Segmentation and CNN Classification with Grad-CAM Explainability

• Marcelo Colares da Silva IFCE
• Caio Marques Silva IFCE
• Suane Pires P. da Silva IFCE
• Roger M. Sarmento IFCE
• Houbing Hebert Song UMBC
• Pedro Pedrosa Rebouças Filho IFCE

Resumo

Diabetic retinopathy (DR) is a diabetes-related ocular complication that can lead to vision loss. Its detection is performed through fundus examinations, assisted by lesion segmentation techniques. The IDRiD and APTOS-2019 datasets are used for DR lesion segmentation and classification, respectively. Using the U-Net architecture, lesions such as microaneurysms and exudates were segmented, while CNNs classified disease stages. In this paper, we present DR-AIVis, an approach for segmentation, classification, and explainability of diabetic retinopathy (DR). Our results demonstrate an accuracy of 93.84% in segmentation and 98.30% in classification. Additionally, we employ Grad-CAM to highlight the most relevant regions of the image. As contributions, our work includes an automated system for DR segmentation and classification, as well as a mechanism to identify the most important image regions for decision-making, thereby enhancing confidence in the provided results using Grad-CAM.