Classification of Severity of Lung Parenchyma Using Saliency and Discrete Cosine Transform Energy in Computed Tomography of Patients With COVID-19

Abstract

This study proposes an automated system for assessing lung damage severity in coronavirus disease 2019 (COVID-19) patients using computed tomography (CT) images. These preprocessed CT images identify the extent of pulmonary parenchyma (PP) and ground-glass opacity and pulmonary infiltrates (GGO-PIs). Two types of images-saliency (Q) image and discrete cosine transform (DCT) energy image-were generated from these images. A final fused (FF) image combining Q and DCT of PP and GGO-PI images was then obtained. Five convolutional neural networks (CNNs) and five classic classification techniques, trained using FF and grayscale PP images, were tested. Our study is aimed at showing that a CNN model, with preprocessed images as input, has significant advantages over grayscale images. Previous work in this field primarily focused on grayscale images, which presented some limitations. This paper demonstrates how optimal results can be obtained by using the FF image rather than just the grayscale PP image. As a result, CNN models outperformed traditional artificial intelligence classification techniques. Of these, Vgg16Net performed best, delivering top-tier classification results for COVID-19 severity assessment, with a recall rate of 95.38%, precision of 96%, accuracy of 95.84%, and area under the receiver operating characteristic (AUROC) curve of 0.9585; in addition, the Vgg16Net delivers the lowest false negative (FN) results. The dataset, comprising 44 COVID-19 patients, was split equally, with half used for training and half for testing.

Keywords: COVID-19; SARS-CoV-2 virus; artificial intelligence; computed tomography; medical diagnostic imaging.

Figure 1

Computed tomography of the chest of a patient with nonserious COVID-19. (a) Without processing. (b) With processing. CT scan of a severe COVID-19 patient. (c) Without processing. (d) With processing.

Figure 2

(a) Computed tomography of the chest of a patient with serious COVID-19. (b) Superpixel clusters for feature extraction. (c) Segmentation and identification of lung PP. (d) Regrouping and identification of the GGO-PI regions.

Figure 3

Block diagram of the implemented code.

Figure 4

Overall method description for COVID-19 CT image severity classification.

Figure 5

Fusion image results of Q and DCT. (a–d) Grayscale PP without severity image results without processing. (e–h) FF image result in processing. (i–l) Grayscale PP with severity image results without processing. (m–p) FF image result in processing.