A Lightweight and Erudite Automated BiLSTM ‐Based Multi‐Cycle GAN for Retinal Blood Vessel and Optic Disc Segmentation in Fundus Images

Accurate segmentation of retinal blood vessels and optic discs is vital for the early diagnosis of ophthalmic conditions such as diabetic retinopathy and glaucoma. However, existing methods often struggle with computational complexity and poor generalization across diverse datasets. To address these challenges, we propose a novel Bidirectional Long Short‐Term Memory (BiLSTM)‐based Multi‐Cycle Generative Adversarial Network (MCGAN). This architecture integrates the temporal modeling strength of BiLSTM with the adversarial learning capacity of a cycle‐consistent GAN to improve the spatial coherence and accuracy of thin and thick vessel and optic disc segmentation in fundus images. The BiLSTM module captures spatial‐sequential dependencies in retinal structures by encoding image data as ordered sequences, enhancing the continuity of thin and thick vessels and the delineation of anatomical boundaries. These extracted features are passed to MCGAN, which refines and reconstructs the segmentation mask. The discriminator network enforces structural realism by distinguishing between ground truth and generated segmentations. A multi‐cycle consistency loss ensures the preservation of anatomical integrity through reconstruction of the original fundus image from the predicted mask, even under weak supervision. An attention‐guided refinement module further improves focus on clinically relevant regions. The model is optimized using a composite loss function combining dice loss, focal loss, adversarial loss, and structural similarity measures to handle class imbalance and enforce semantic consistency. Extensive evaluations on publicly available datasets DRIVE, and STARE demonstrate that the proposed model achieves superior segmentation performance with reduced computational complexity. The proposed model achieves an average accuracy of 99.12%, recall of 98.78%, precision of 98.22%, f1‐score of 97.13%, and JC of 96.67% on the DRIVE dataset, and 97.80% accuracy, 98.36% precision, f1‐score of 98.74%, recall of 97.74%, and JC of 95.80% on the STARE dataset. The proposed framework offers robust, generalizable, and efficient segmentation capabilities, making it highly suitable for real‐world ophthalmic diagnostic systems.