Enhancing fundus image analysis for diabetic retinopathy using CheXNet with CBAM and Grad-CAM visualization

TL;DR

This paper introduces a deep learning model for detecting diabetic retinopathy in fundus images, achieving high accuracy with attention mechanisms and visualization tools.

Contribution

The novel integration of CBAM with CheXNet-initialized DenseNet121 for DR classification, along with Grad-CAM for interpretability.

Findings

The CheXNet_CBAM model achieved 96.12% accuracy on the APTOS 2019 dataset.

It also attained 96.33% accuracy on the DDR dataset, outperforming other architectures.

CBAM improved feature learning in the DenseNet121 framework for DR classification.

Abstract

Diabetic retinopathy (DR) is a leading cause of vision impairment among individuals with diabetes. Early detection and accurate grading are essential for timely clinical management. However, developing robust models for automated interpretation and grading of fundus images remains challenging due to variability in lesion appearance and image quality.

This study proposes a deep learning framework for DR classification from fundus images based on a DenseNet121 backbone initialized with CheXNet weights. A Convolutional Block Attention Module (CBAM) is integrated to enhance feature representation through channel and spatial attention mechanisms in a data-driven manner. In addition, Gradient–weighted Class Activation Mapping (Grad–CAM) is employed to provide post hoc visual explanations of model predictions. The proposed CheXNet_CBAM model is evaluated against several convolutional neural network architectures, including CheXNet, DenseNet121, MobileNetV2, VGG19, and ResNet50, using the APTOS 2019 and DDR datasets.

On the APTOS 2019 dataset, the proposed model achieves an accuracy of 96.12%, while on the DDR dataset it attains 96.33%, outperforming the compared architectures on both benchmarks.

The results indicate that incorporating CBAM improves discriminative feature learning within a DenseNet121–based framework. While the model demonstrates strong performance across two public datasets, further prospective evaluation and external validation are required to assess its clinical applicability in real–world settings.

Figures