Att-SV2-ES Net: Hybrid deep learning features and lightweight attention-based model for early glaucoma detection

Glaucoma is one of the main causes of permanent blindness. Timely intervention is ensured by its early and precise identification. This paper proposes a new hybrid approach, integrating handcrafted features with Deep Learning (DL), for reliable glaucoma identification and prediction on the G1020 dataset. Initially, pre-processing is conducted, which includes data augmentation to increase the diversity of images, Gaussian filtering to reduce image noise, Contrast Limited Adaptive Histogram Equalization (CLAHE) to enhance local retinal contrast and then grayscale conversion to simplify retinal image representation. After pre-processing, the enhanced images were given to DDSAM (Depth-wise dilated Spatial Attention mechanism) based Glau-SegNet for optic disc (OD) and optic cup (OC) segmentation. From the segmented images, handcrafted features such as Local Binary Patterns (LBPs) and fractal analysis, which describe complementary texture and structural information, are extracted. In parallel, the segmented images are processed by the Attention ShuffleNet V2 with Echo State Network (Att-SV2-ES Net) to obtain high-level deep feature representations. Then these deep features and handcrafted features are concatenated and fed to a fully connected layer with sigmoid activation to classify into the class for the prediction of glaucoma. This combined approach achieves a balance between high accuracy and scalability, addressing the growing need for accessible and reliable glaucoma screening solutions. The experimental results demonstrate the proposed system's superior performance in detecting and predicting glaucoma. Furthermore, the findings demonstrate that the proposed Model, which integrates DL features, outperforms all other models in terms of key performance metrics, including achieving 99.32% specificity, 99.43% sensitivity, and 99.37% accuracy.