Fused framework for glaucoma diagnosis using Optical Coherence Tomography (OCT) images

Glaucoma is the second leading cause of vision loss induced by impairment of the optic nerve. Manual procedure for diagnosing glaucoma is associated with several challenges such as human error and time consumption, due to which an alternative technology assisted diagnosis methodology is preferable. The preponderance of research involves machine learning or deep learning for classification of glaucoma, and employs the commonly used imaging modalities, fundus imaging and Optical Coherence Tomography (OCT) for classification of an eye as Normal or Glaucomatous. A fused framework has been proposed for the same, which involves fusion of predicted classes from machine learning and deep learning techniques using Major Voting and Weighted Decision Fusion. This research work further explores the variability in predicted classes and identifies an optimum prediction threshold for comparison with existing techniques. Based on experimentation performed, a Precision of 0.95, Recall of 0.97 and F1-Score of 0.96 have been observed for the proposed Major Voting Fusion Framework, which outperformed k-Nearest Neighbours (0.93, 0.97, 0.95), Random Forest Classifier (0.94, 0.92, 0.93), Support Vector Machines (0.92,0.94,0.93) and 3D-Convolutional Neural Network (0.91, 0.91, 0.91) based deep learning approach. Gradient Weighted Class Activation Maps (Grad-CAMs) are also plotted to assist medical experts in diagnosis. The proposed framework could be beneficial for examining glaucomatous patients using OCT images. The higher F1 Score also indicates robustness of our approach on an uneven number of normal and abnormal samples in the dataset as well.

Automated summary

This research investigates a fused framework for assisting in the diagnosis of glaucoma using machine learning and deep learning techniques, presenting a method of fusing predicted classes through Major Voting and Weighted Decision Fusion. Experimental results demonstrate that this framework outperforms deep learning approaches based on other models, with a Precision of 0.95, Recall of 0.97 and F1-Score of 0.96, while also providing Grad-CAMs to aid medical experts in diagnosis.