A fast lightweight network for the discrimination of COVID-19 and pulmonary diseases. (September 2022)

• Record Type: Journal Article
• Authors: Aiadi, Oussama; Khaldi, Belal
• Is Part Of: Biomedical signal processing and control. Volume 78(2022)
• Publication Date: 2022-09
• Subjects: COVID-19 -- Pulmonary diseases -- Chest X-Ray -- Deep learning -- Pneumonia -- Features learning -- Telemedicine; 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.2022.103925
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: We propose a computationally fast network (DLNet) for recognizing COVID-19 and pulmonary diseases. DLNet can be used in telemedicine environment because it has a high tolerance to missed parts in the medical image. DLNet jointly encodes local binary patterns along with feature maps produced by the convolution layer. Convolution is performed using Discrete cosine transform (DCT) filters. We conduct extensive experiments on a public dataset. DLNet has been shown to be effective, computationally fast and robust against missed parts of the medical image. Abstract: With the outbreak of COVID-19 and the increasing number of infections worldwide, there has been a noticeable deficiency in healthcare provided by medical professionals. To cope with this situation, computational methods can be used in different steps of COVID-19 handling. The first step is to accurately and rapidly diagnose infected persons, because the time taken for the diagnosis is among the crucial factors to save human lives. This paper proposes a computationally fast network for the diagnosis of COVID-19 and pulmonary diseases, which can be used in telemedicine. The proposed network is called DLNet because it jointly encodes local binary patterns along with filter outputs of discrete cosine transform (DCT). The first layer in DLNet is the convolution layer in which the input image is convolved using DCT filters. Then, to avoid over-fitting, a binary hashing procedure is performed by fusing responses of different filters into a unique feature map. This map is used to generate block-wise histograms by binding local binary codes of the input image and the map values. We normalize these histograms to improve the robustness of the network against illumination changes. Experiments conducted on a public dataset demonstrate the rapidity and effectiveness of DLNet, where an average accuracy, sensitivity, and specificity of 98.86%, 98.06, and 99.24% have been achieved, respectively. Moreover, the proposed network has shown high tolerance to the missing parts in the medical image, which makes it suitable for the telemedicine scenario.