# Deep Learning Approach to Diabetic Retinopathy Detection

**Authors:** Borys Tymchenko, Philip Marchenko, Dmitry Spodarets  
**arXiv:** [2003.02261](https://arxiv.org/abs/2003.02261)  
**Date:** 2020-03-03  
**Summary page:** https://www.summarizepaper.com/en/paper/2003.02261v1/

## Key points

- Early detection in diabetic retinopathy is crucial to prevent permanent blindness
- Identification of disease stage is challenging and requires expert human interpretation of fundus images
- Authors propose an automatic deep-learning-based method using a single photograph of the human fundus for stage detection
- Method utilizes convolutional neural networks (CNN) for diagnosing diabetic retinopathy
- High cost of labeled datasets and inconsistencies between doctors hinder performance of these methods
- Authors propose a multistage approach to transfer learning, leveraging similar datasets with different labeling
- Method achieves sensitivity and specificity of 0.99 and ranked 54 out of 2943 competing methods on APTOS 2019 Blindness Detection Dataset
- SHAP is used to visualize features contributing to disease stage assessment, ensuring useful feature learning during training
- This deep learning approach has potential to automate detection and diagnosis, improving accuracy in early detection

## Summary

The paper titled "Deep Learning Approach to Diabetic Retinopathy Detection" discusses the importance of early detection in diabetic retinopathy, a complication of diabetes that can lead to permanent blindness if left untreated. The exact identification of the disease stage is challenging and requires expert human interpretation of fundus images. To simplify this process, the authors propose an automatic deep-learning-based method for stage detection using a single photograph of the human fundus. The method utilizes convolutional neural networks (CNN), which have been successful in diagnosing diabetic retinopathy and other related subjects. However, the performance of these methods is hindered by the high cost of labeled datasets and inconsistencies between different doctors. To address these issues, the authors propose a multistage approach to transfer learning, which leverages similar datasets with different labeling. The presented method demonstrates promising results as a screening tool for early detection of diabetic retinopathy. It achieves a sensitivity and specificity of 0.99 and is ranked 54 out of 2943 competing methods on the APTOS 2019 Blindness Detection Dataset, with a quadratic weighted kappa score of 0.925466. In addition to evaluating binary classification (DR/No DR) for assessing model quality as a screening method, the authors also discuss interpreting models' predictions in medical applications. They use SHAP (Shapley Additive exPlanations) to visualize features that contribute to disease stage assessment, ensuring that the model learns useful features during training. Overall, this paper presents a deep learning approach that shows promise in automating the detection and diagnosis of diabetic retinopathy. By simplifying and improving accuracy in early detection, this method has significant potential to help millions of people at risk for permanent blindness due to this complication of diabetes.