A unified deep learning framework for multi-lesion detection in diabetic retinopathy using fundus images
Year: 2026
Type: article
Abstract: Diabetic Retinopathy (DR), an ocular disease caused by diabetes, is a significant cause of vision impairment. Early and accurate detection of different types of retinal lesions, including Microaneurysms, Haemorrhages, Soft Exudates and Hard Exudates, is essential for DR diagnosis and prevention of associated vision loss. Conventional manual analysis of retinal fundus images is subjective and time-consuming, requiring specialized clinical expertise. Automated diagnostic systems can alleviate these issues and assist clinicians with objective analysis and improved outcomes. This work focuses on simultaneous detection of different lesion types in DR, through a unified framework. In particular, the U-Net architecture is employed for multi-lesion segmentation due to its ability to preserve fine structural details using its encoder-decoder design. Additionally, an Adaptive Batch Scaling (ABS) strategy is employed, wherein the batch size is dynamically adjusted across training epochs, based on the validation Dice score, to optimize pixel-level segmentation accuracy. The performance of U-Net is compared with other DL-based segmentation models such as Mask Region-based Convolutional Neural Network, YOLOv8, and Faster Region-based Convolutional Neural Network architectures, in terms of its ability to accurately segment and identify diverse DR lesions in retinal fundus images. Validation is performed on two different publicly available DR datasets, i.e., Indian Diabetic Retinopathy Image Dataset (IDRiD) and Diabetic Retinopathy Detection Dataset (DDR). Specifically, DDR dataset is used for training the model, which is tested separately on each of the two datasets – DDR and IDRiD, to assess intra-dataset and cross-dataset generalization. Performance metrics including accuracy, Dice score, Intersection over Union (IoU) and mean average precision (mAP), are used to evaluate the performance of the models. The findings reveal that the U-Net architecture demonstrates superior performance, achieving an accuracy of 0.925, Dice score of 0.771, IoU of 0.748 and mAP of 0.647. The results demonstrate the potential of semantic segmentation models, particularly U-Net, in automated DR screening. The proposed unified multi-lesion detection system enables simultaneous detection of multiple DR lesions in a single-pass reducing computational overhead, while maintaining reasonable detection performance. The system is envisaged to serve as a reliable tool in DR screening workflow, facilitating early diagnosis and timely intervention, thereby reducing the burden of DR-related vision worldwide.
Source: Scientific Reports
Authors: Remya Ramesh, R. Lavanya, O. K. Sikha
Institutions: Amrita Vishwa Vidyapeetham, Universitat Pompeu Fabra
Language: English
Topic: Retinal Imaging and Analysis
Subfield: Radiology, Nuclear Medicine and Imaging
Field: Medicine
Domain: Health Sciences
Open Access status: gold
Estimated APC paid: $2,690
OpenAlex ID: W7222924592