# Introducing Feature Attention Module on Convolutional Neural Network for Diabetic Retinopathy Detection

**Authors:** Susmita Ghosh, Abhiroop Chatterjee  
**arXiv:** [2308.02985](https://arxiv.org/abs/2308.02985)  
**Date:** 2023-08-06  
**Summary page:** https://www.summarizepaper.com/en/paper/2308.02985v1/

## Key points

- Deep learning models have shown promising results in automating the process of diabetic retinopathy (DR) detection.
- The study proposes a new methodology that integrates a feature attention module with a pretrained VGG19 convolutional neural network (CNN) to improve DR detection accuracy.
- The feature attention module selectively highlights salient features from fundus images and fuses them with the original input, allowing for simultaneous learning of attention weights and combination of attention-modulated features.
- The proposed method was evaluated using the APTOS DR Dataset and achieved an impressive accuracy rate of 95.70%.
- Integrating the feature attention module with a pretrained CNN effectively enhances DR detection accuracy by leveraging complementary information and selectively highlighting salient features.
- These findings contribute to developing automated systems for early detection and diagnosis of diabetic retinopathy, leading to improved patient outcomes and reduced blindness among diabetic individuals.

## Summary

In the field of diabetic retinopathy (DR) detection, deep learning models have shown promising results in automating the process. In this study, authors Susmita Ghosh and Abhiroop Chatterjee propose a new methodology that integrates a feature attention module with a pretrained VGG19 convolutional neural network (CNN) to improve the accuracy of DR detection. The pretrained net is fine-tuned with the proposed feature attention block which aims to leverage complementary information from various regions of fundus images to enhance the discriminative power of the CNN. The feature attention module incorporates an attention mechanism that selectively highlights salient features from images and fuses them with the original input. This allows for simultaneous learning of attention weights for the features and combination of attention-modulated features within the feature attention block. By doing so, the network is able to focus on relevant information while reducing the impact of noisy or irrelevant features. To evaluate their proposed method, the authors used a widely used dataset for diabetic retinopathy classification called APTOS (Asia Pacific Tele-Ophthalmology Society) DR Dataset. They compared their results with and without the attention module as well as with other state-of-the-art approaches. The results confirmed that introducing the fusion module (fusing of feature attention module with CNN) significantly improved accuracy rate achieving an impressive accuracy rate of 95.70%. Overall, this study demonstrates that integrating a feature attention module with a pretrained CNN can effectively enhance DR detection accuracy by leveraging complementary information and selectively highlighting salient features from fundus images. These findings contribute to ongoing efforts in developing automated systems for early detection and diagnosis of diabetic retinopathy ultimately leading to improved patient outcomes and reduced blindness among diabetic individuals.