Effective Identification of Glaucoma using Fusion Deep Learning Approaches

• Published in: 2024 International Conference on Electronic Systems and Intelligent Computing (ICESIC) pp. 7 - 11
• Main Authors: J, Manikandan, S, Karthik, L, Naveen
• Format: Conference Proceeding
• Language: English
• Published: IEEE 22.11.2024
• Subjects: Accuracy; AlexNet; Blindness; Computer architecture; Convolutional neural networks; Dense-Net; Feature extraction; Fivefold cross validation; Glaucoma; Heuristic algorithms; Robustness; Routing; Sensitivity
• DOI: 10.1109/ICESIC61777.2024.10846790

Glaucoma has emerged as one of the main reasons for irreversible blindness in the world. Thus, there is an urgent need for diagnostic techniques that are accurate and timely. In this work, the method proposes using five-fold cross-validation over the combination of AlexNet and DenseNet-201 for Glaucoma classification using the ONH and ODT datasets. AlexNet is designed based on the basis of deep convolutional layers to extract significant features, while DenseNet-201 maintains the spatial hierarchies of the image data, thus preventing the loss that usually occurs through convolutional neural networks. Extensive pre-processing of the OCT scans was done to improve the quality and relevance of the input data. Pre-processing involved the pipeline of image normalization and augmentation techniques, such as flip and zoom, so as to make the model robust against variations in image acquisition. It uses AlexNet for extracting the initial feature representations from OCT scans, further refining the spatial relationship of those features using DenseNet-201's dynamic routing algorithm. This is a dual-model architecture based on the hypothesis that such a structure can combine the complementary strengths of both networks, therefore increasing the classification accuracy and robustness in distinguishing benign from malignant Eye nodules. Our model was tested on metrics such as accuracy, sensitivity, and F1 score, which performed considerably better than the existing methodologies. Experimental results on the datasets of ONH and OCT are indicative that our hybrid model achieves 97.8% in terms of accuracy, higher than that of the present top methods. The capability of the new model to preserve the spatial structure and enhance feature extraction is an important step toward medical diagnosis that could be improved in clinical scenarios.