Convolution neural networks for optical coherence tomography (OCT) image classification. (January 2023)

• Record Type: Journal Article
• Authors: Karthik, Karri; Mahadevappa, Manjunatha
• Is Part Of: Biomedical signal processing and control. Volume 79(2023)Part 2
• Publication Date: 2023-01
• Subjects: Image classification -- Optical coherence tomography (OCT) -- Convolution neural networks -- Retinal diseases -- Cross activation; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2022.104176
• Languages: English
• ISSNs: 1746-8094

Abstract

Abstract: Optical coherence tomography (OCT) is an imaging modality used to obtain a cross-sectional image of the retina for retinal disease diagnosis. Modern diagnosis systems use Convolutional Neural Networks. Our model increases the contrast in the residual connection, so high contrast regions, such as the retinal layers, are prominent in feature maps. Our model increases the contrast of the derivatives to generate sharper feature maps. We replaced the residual connection in standard ResNet architectures with our design. The proposed activation function retains negative weights and reinforces smaller gradients. We have used two OCT datasets with four and eight classes of diseases, respectively. We performed graphical analysis using Precision–Recall curves. We used accuracy, precision, recall, and F1 score for evaluation. In our laboratory conditions, We have successfully increased the classification accuracy with our proposed design. The gain in accuracy is limited, i.e. < 1% when the initial accuracy is more than 98%, and 1.6% when the initial accuracy is lower. In confusion matrices, we observed the maximum performance increase when the number of samples is less in one class, which will be helpful if data is imbalanced. The retinal boundary is enhanced, with the background (the region outside the retinal layers) suppressed but not entirely removed. In ablation studies, We observed an average accuracy loss of 0.875% with OCT-C4 data and 1.39% for OCT-C8 data. The p-values from Wilcoxon signed-rank test range from 1.65 × 10 −6 to 0.025, and 0.51 for ResNet50 with the OCT-C8 dataset. Graphical abstract: Highlights: Our paper presents a residual module designed for OCT image classification for retinal diseases. Our methodology enhances the contrast of feature maps, resulting in more distinct retinal layer boundaries. We propose an activation function to retain negative weights and boost smaller gradients at the final fully connected layer. Our activation function introduces non-linearity in gradients, with a differential rate of change based on the current value of the gradient.