A federated learning framework integrating an improved bidirectional long short-term memory architecture has achieved high accuracy in detecting diabetic retinopathy and estimating condition severity across image test sets. The system, designated as the FLB-DRD-SE model, demonstrated an accuracy of 98.70% on both experimental datasets evaluated in the study.
Diabetic retinopathy is a widespread complication of diabetes mellitus that produces lesions on the retina, impairing vision and potentially resulting in blindness if not identified early. While deep learning algorithms have become important tools that deliver improved diagnostic performance, experimental metrics and corresponding datasets in existing studies continue to suffer from imbalance and bias.
To address these challenges, the investigators established a pipeline that begins with image enhancement and anatomical segmentation on retinal fundus inputs. The workflow first applies the Wiener filtering method to pre-process fundus photographs, followed by image segmentation executed through a specialized BNMSK-U-Net model.
Following image segmentation, the computational framework extracts multiple representations to capture diagnostic indicators. These derived properties include deep features, pyramid histogram of oriented gradients (PHOG)-based features, and Improved Local Gabor Increasing Pattern (ILGIP) features extracted directly from the segmented fundus photographs.
The decentralized training mechanism relies on a Modified Activation and Loss function-assisted Bi-directional Long Short-Term Memory (MAL-Bi-LSTM) network running locally on individual client devices.
Applying the clustered global model alongside the extracted visual features enabled the system to execute both diabetic retinopathy detection and severity estimation. In performance evaluations, the FLB-DRD-SE pipeline achieved 98.70% accuracy, 99.00% precision, and 97.30% F-measure on Dataset 1, while registering 98.70% accuracy, 84.30% precision, and 86.00% F-measure on Dataset 2.
Subject of Research: Diabetic retinopathy detection and severity estimation using federated learning
Article Title: Federated Learning-based Improved Bi-LSTM Network Model for Diabetic Retinopathy Detection and Severity Estimation
Article References: Nair, P. G., & Lanitha, B. (2026). Federated Learning-based Improved Bi-LSTM Network Model for Diabetic Retinopathy Detection and Severity Estimation. International Journal of Computational Intelligence Systems. https://doi.org/10.1007/s44196-026-01647-8
Image Credits: AI Generated
DOI: 10.1007/s44196-026-01647-8
Keywords: diabetic retinopathy, federated learning, Bi-LSTM, severity estimation, fundus images, Wiener filtering, BNMSK-U-Net, PHOG features, ILGIP features, deep features, K-means clustering, MAL-Bi-LSTM, FLB-DRD-SE, retinal lesions, diabetes mellitus, global aggregation model, local model training
News Source: Veronica Carney. (November 20, 2026). Federated Learning Bi-LSTM Model Detects Diabetic Retinopathy and Estimates Severity. Scienmag.



