Gradient-Weighted Class Activation Mapping Based Deep Transfer Learning For Glaucoma Disease Prediction

Glaucoma is a progressive eye disease characterized by damage to optic nerve. Early detection and management are crucial to preserving vision, making prediction of glaucoma risk. To improve accurate prediction, Gradient-weighted Class Activation Mapped Deep Transfer Learning (GWCAMDTL) model is developed. The main aim of the model is to enhance the accuracy of glaucoma prediction while minimizing time consumption. Retinal fundus images are collected from the dataset for accurate prediction in image acquisition phase. Deep transfer learning involves adapting a pre-trained deep learning model for performing glaucoma prediction. In proposed Deep transfer learning model, the Multilayer Perceptron classifier is used as pre-trained model for analyzing the given large number of training images. Then, new model is constructed along with its pre-trained model for disease prediction. Initially, layers in pre-trained model are usually frozen to preserve the learned features from the infected regions. Transferring information from previously learned results by the pre-trained mode to new tasks has the potential to significantly improve feature learning efficiency by applying the congruence correlation coefficient. Gradient-weighted Class Activation Mapping generates visual explanations for the predictions made by model. Fine-tuning layers is crucial part of transfer learning. During fine-tuning for glaucoma prediction, the model weights of certain layers are updated to better fit the specific characteristics of the new glaucoma dataset, leading to a reduction in both training and validation error. This approach improves the accuracy of glaucoma prediction by applying the strengths of the pre-trained model and adapting it to the clinical features of retinal fundus images. This process helps to make accurate predictions and extensively improves the F1-score. Experimental are conducted using various evaluation metrics. Results of GWCAMDTL achieve higher accuracy with reduced time as well as error compared to existing methods.