Redundancy reduced depthwise separable convolution for glaucoma classification using OCT images

Glaucoma is a chronic progressive optic neuropathy characterized by the impairment of the optic nerve, and if it is left untreated, it may lead to irreversible vision loss. So, the early detection, diagnosis, and management of retinal glaucoma are necessary as the number of cases increases expeditiously. Existing deep learning classification methods of glaucoma are computationally intensive, restricting memory potency, training, and affects the optimization of hyperparameters. Thus they are unsuitable for real-time applications with limited computing resources. A two-dimensional depthwise separable convolution architecture can significantly improve the efficiency of parameter utilization and calculation speed. This paper proposes a raw SD-OCT-based depthwise separable convolution model to classify glaucoma from healthy images. Each input channel is convolved with each filter kernel, and the resulting output channels are effectively mixed by pointwise convolution. A gradientweighted class activation mapping is estimated to highlight the region with structural deformations per B scan to validate the performance qualitatively. Our private database comprises 1105 glaucomatous and 1049 normal OCT B scans around ONH, which aids the ophthalmologists in making an appropriate diagnosis. The proposed redundancy reduced depthwise separable convolution network achieved an accuracy, precision, recall, F1-score, and AUC of 0.9963, 0.9946, 0.9982, 0.9964, and 0.9963, respectively. The number of parameters used in the proposed method is only 20,686. In the proposed network, tedious segmentation and other image processing steps are avoided, still achieving remarkable network training efficacy with appreciable reduction of learnable parameters.