Fusion-CapsNet: A transfer-driven capsule network trained on fused features and generative ai augmentation for...

Strabismus is a common eye condition in which the eyes are does not line up correctly that impacts peoples of all ages. For effective treatment and to prevent durable problems, early detection and accurate classification are essential. This research presents a Transfer-driven Capsule Network trained on Fused features and generative AI augmentation method for strabismus detection, named as Fusion-CapsNet. Firstly, an improved Wasserstein Generative Adversarial Network with Gradient Penalty (ImWGAN-GP) method is utilized to generate high-quality synthetic eye images, which improves enhance the data diversity and model generalization. Following image augmentation, a Medav (median-average) based preprocessing is utilized to avoid the noise while preserving the critical structural as well as edge information. Then, accurate localization of visual regions is obtained by using an improved TransUNet (ImTransUNet) segmentation model. It integrates improved activation functions and an Atrous Spatial Pyramid Pooling-Progressive Enhancement Module (AS-IPEM) to effectively capture both local spatial details and global contextual information. From the segmented eye regions, multiple complementary features are extracted to widely characterize the strabismic patterns such as Eye Aspect Ratio (EAR), Local Gabor Transitional Patterns (LGTrP) and Active Shape Models (ASM). An improved weighted PCA with adaptive concatenation and dynamic scaling (ImWPCA-ACDS) based feature fusion strategy is introduced to effectively integrate these heterogeneous features. It preserves the semantic consistency, captures long-range dependencies, reduces feature dimensionality and mitigates overfitting. Lastly, strabismus detection is performed via an improved CapsNet with transfer learning (ImCapsNet-TL), which enables the model to effectively capture the hierarchical spatial relationships and long-range feature interactions while reducing the computational overhead. Experimental outcomes validate that the proposed Fusion-CapsNet framework obtained superior detection performance compared to standard methods. The ImCapsNet-TL detection frameworks has resulted higher accuracy by 97.7% in data size of 90%.