Multi-classification of fundus diseases based on DSRA-CNN. (August 2022)

• Record Type: Journal Article
• Authors: Yang, Xue-lian; Yi, San-li
• Is Part Of: Biomedical signal processing and control. Volume 77(2022)
• Publication Date: 2022-08
• Subjects: Fundus diseases -- Depthwise separable convolutions -- Residual connections -- SE block; 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.2022.103763
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: A new model realizes eight fundus disease diagnosis. The model includes pre-processing of input data, feature extraction network, and classifiers. A feature extraction network named DSRA-CNN is proposed. We achieve better results than advanced convolutional neural networks. Abstract: There are many kinds of fundus diseases, and early diagnosis is the key to prevent severe visual impairment. In this paper, we propose a deep learning model to automatically detect multiple fundus diseases which includes three parts. The first part is the pre-processing of the data including data screening, black border cropping, data augmentation, and normalization. The second part is a feature extraction network named DSRA-CNN based on Xception architecture, which combines the function blocks of DS block, DSR block, and SE block, among which DS block is composed of depthwise separable convolutions to reduce the amount of calculation, DSR block is composed of DS blocks and residual connections to strengthen the utilization of image features, SE block is used to screen the characteristic information. Lastly, based on the extracted features, we design a classifier to realize the classification of eight different fundus diseases. Experiments are performed on the ODIR dataset and the results of the accuracy, precision, F1 value, and kappa score of the DSRA-CNN network proposed in this paper are respectively 87.90%, 88.50%, 88.16%, and 86.17%. Compared with the original Xception network and existing advanced Convolutional Neural Networks(CNNs), our proposed network performs better.