Segmentation-Driven Morphology-Guided Attention Mechanism for Rhegmatogenous Retinal Detachment Classification

Rhegmatogenous Retinal Detachment (RRD) is a vision-threatening condition that requires accurate classification to prevent vision loss. The existing Deep Learning (DL) models for RRD primarily rely on generic feature extraction from fundus images and often overlook subtle morphological variations across RD subtypes. To address this issue, an effective integrated model was proposed by combining EfficientNetB3-based DeepLabV3+ segmentation with morphological feature extraction and a morphological-guided attention mechanism referred to as Morpho-RRDNet. The segmentation stage isolates detachment zones by enabling RDD-relevant morphometric descriptors, including area percentage, centroid location, and convexity. These features were fused with deep spatial embeddings using a Morphological-Guided Attention Mechanism (MGAM) to focus mainly on RRD-specific regions while suppressing irrelevant regions. Finally, a Light Gradient Boosting Model (LightGBM) was introduced to classify the RRD and non-RRD classes. The proposed Morpho-RRDNet achieved competitive results, including accuracy (99.22%), sensitivity (100%), specificity (99.20%), and F1-score (85.71%), when compared to the existing optimized Ensemble Network.