Hyper‐reflective foci segmentation in SD‐OCT retinal images with diabetic retinopathy using deep convolutional neural networks

Abstract

PURPOSE: The purpose of this study was to automatically and accurately segment hyper-reflective foci (HRF) in spectral domain optical coherence tomography (SD-OCT) images with diabetic retinopathy (DR) using deep convolutional neural networks. METHODS: An automatic HRF segmentation model for SD-OCT images based on deep networks was constructed. The model segmented small lesions through pixel-wise predictions based on small image patches. We used an approach for discriminative features extraction for small patches by introducing small kernels and strides in convolutional and pooling layers, which was applied on the state-of-the-art deep classification networks (GoogLeNet and ResNet). The features extracted by the adapted deep networks were fed into a softmax layer to produce the probabilities of HRF. We trained different models on a dataset with 16 HRF eyes by using different sizes of patches, and then, we fused these models to generate optimal results. RESULTS: Experimental results on 18 eyes demonstrated that our method is effective for the HRF segmentation. The dice similarity coefficient (DSC) for the foci area in B-scan, projection images, and foci amount in B-scan images reaches 67.81%, 74.09%, and 72.45%, respectively. CONCLUSIONS: The proposed segmentation model can accurately segment HRF in SD-OCT images with DR and outperforms traditional methods. Our model may provide reliable segmentations for small lesions in SD-OCT images and may be helpful in the clinical diagnosis of diseases.