Hybrid Framework for Explainable Glaucoma Detection: Integration of U-Net, DeepLabV3+, and Semi-Supervised Learning

In this paper, we present an image based hybrid methodology which is combination of convolution methodologies for diagnosing glaucoma via retinal fundus images. It combines U-Net and DeepLabV3+ for multiscale segmentation with strong consistency, and a semi-supervised classifier involving Shepard Convolutional Neural Network (ShCNN) and Deep Kronecker Network (DKN) for accurate and robust classification in the presence of labeled and pseudo-labeled data. The hybrid combination obtains better generalization and robustness rather than individual models. It remedies a number of generic challenges, including dataset noise, scarce labels, and clinical explainability. The proposed framework reaches the accuracy, sensitivity and specificity of 98.7%, 98.2% and 99.1% respectively, performing superior than the original models of CNNs and SVMs. Explainable AI methods such as Grad-CAM also offer clinical interpretability to be safe for routine use.