FRCNN: A Combination of Fuzzy-Rough-Set-Based Feature Discretization and Convolutional Neural Network for Segmenting Subretinal Fluid Lesions

Abstract

Segmentation of subretinal fluid (SRF) lesions in spectral domain optical coherence tomography (SD-OCT) images is essential in the quantitative analysis of fundus lesion diagnosis. Owing to noise interference as well as differences in lesions between patients, the precise segmentation of SRF lesions is extremely difficult. To this end, a segmentation method (FRCNN) of SRF lesions, which combines fuzzy-rough-set-based feature discretization and convolutional neural network is proposed. First, each segmented region category is regarded as a fuzzy cluster, the fuzzy clustering is employed to derive the membership functions of all segmented region categories according to the number of fuzzy clusters, and the fuzzy relationship is determined by calculating the distance between pixels, thus accurately describing the sample blurriness in SD-OCT images. Second, pixel values are discretized by building a fuzzy-rough-set-based fitness function in a rough approximation space for the redundant information and noise in the image. This function is composed of the number of breakpoints and the average approximation precision of fuzzy rough sets of all segmented region categories, which can reduce the data size while ensuring high accuracy. Finally, the images after feature discretization are taken as input data, and the deep attention modules with hybrid kernel convolutions are used to capture multiscale information in the fully convolutional neural network architecture. Compared with advanced SD-OCT fundus image segmentation algorithms, FRCNN can achieve better segmentation results and effectively segment the SRF lesions, providing strong support for theoretical research and clinical applications.