Federated Learning-based Improved Bi-LSTM Network Model for Diabetic Retinopathy Detection and Severity Estimation

A common complication of diabetes mellitus is Diabetic Retinopathy (DR), which causes lesions on the retina that affect vision. It can lead to blindness if it is not detected early. Recently, Deep Learning (DL) has become a crucial technique, which has attained better performance in DR detection. However, metrices utilized in the experiments and their respective dataset remain unbalanced and biased across diverse studies. Consequently, a Federated Learning (FL) based Bi-LSTM network for Diabetic Retinopathy Detection and Severity Estimation (FLB-DRD-SE) model is introduced in this work. The Wiener filtering method is first used to pre-process the input fundus image. The pre-processed image is then segmented via the BNMSK-U-Net model. The segmented images are subsequently used to extract PHOG-based features, deep features, and Improved Local Gabor Increasing Pattern (ILGIP) features. With these extracted features, a local model is obtained by the proposed Modified Activation and Loss function-assisted Bi-directional Long Short-Term Memory (MAL-Bi-LSTM) model. Afterwards, this local model is given to central server, which employs the K-means clustering technique, which groups similar updates based on which clients have similar models and provides a global aggregation model as output. Moreover, the FL framework utilizes MAL-Bi-LSTM model for local training; each device performs training on its patient’s data. Finally, using this global model and the extracted features, DR detection and severity estimation are conducted. Moreover, the proposed FLB-DRD-SE model achieved an accuracy, precision, and F-measure of 98.70%, 99.00%, and 97.30%, respectively, on Dataset 1, while it achieved 98.70%, 84.30%, and 86.00%, respectively, on Dataset 2.