SFFT-CapsNet: Stacked Fast Fourier Transform for Retina Optical Coherence Tomography Image Classification using Capsule Network

Abstract

The work of the Ophthalmologist in manually detecting specific eye related disease is challenging especially screening through large volume of dataset. Deep learning models can leverage on medical imaging like the retina Optical Coherence Tomography (OCT) image dataset to help with the classification task. As a result, many solutions have been proposed based on deep learning-based convolutional neural networks (CNNs). However, the limitations such as inability to recognize pose, the pooling operations which affect resolution of the featured maps have affected its performance in achieving the best accuracies. The study proposes a Capsule network (CapsNet) with contrast limited adaptive histogram equalization (CLAHE) and Fast Fourier transform (FFT), a method we called Stacked Fast Fourier Transform-CapsNet (SFFT-CapsNet). The SFFT was used as an enhancement layer to reduce noise in the retina OCT image. A two-block framework of three-layer convolutional capsule network each was designed. The dataset used for this study was presented by University of California San Diego (UCSD). The dataset consists of 84,495 X-Ray images categorized into four classes (NORMAL, CNV, DME, and DRUSEN). Experiment was conducted on the SFFT-CapsNet model and results were compared with baseline models for performance evaluation using accuracy, sensitivity, precision, specificity, and AUC as evaluation metrics. The evaluation results indicate that the proposed model outperformed the baseline model and state-of-the-arts models by achieving the best accuracies of 99.0%, 100%, and 99.8% on overall accuracy (OA), overall sensitivity (OS), and overall precision (OP), respectively. The result shows that the proposed method can be adopted to aid Ophthalmologist in retina disease diagnosis.