Grayscale image statistics of COVID‐19 patient CT scans characterize lung condition with machine and deep learning

Background Grayscale image attributes of computed tomography (CT) of pulmonary scans contain valuable information relating to patients with respiratory ailments. These attributes are used to evaluate the severity of lung conditions of patients confirmed to be with and without COVID‐19. Method Five hundred thirteen CT images relating to 57 patients (49 with COVID‐19; 8 free of COVID‐19) were collected at Namazi Medical Centre (Shiraz, Iran) in 2020 and 2021. Five visual scores (VS: 0, 1, 2, 3, or 4) are clinically assigned to these images with the score increasing with the severity of COVID‐19‐related lung conditions. Eleven deep learning and machine learning techniques (DL/ML) are used to distinguish the VS class based on 12 grayscale image attributes. Results The convolutional neural network achieves 96.49% VS accuracy (18 errors from 513 images) successfully distinguishing VS Classes 0 and 1, outperforming clinicians’ visual inspections. An algorithmic score (AS), involving just five grayscale image attributes, is developed independently of clinicians’ assessments (99.81% AS accuracy; 1 error from 513 images). Conclusion Grayscale CT image attributes can be successfully used to distinguish the severity of COVID‐19 lung damage. The AS technique developed provides a suitable basis for an automated system using ML/DL methods and 12 image attributes. Grayscale CT image statistics accurately distinguish the severity of COVID‐19‐related lung conditions Highlights Grayscale image statistics of CT scans can effectively classify lung abnormalities Graphical trends of grayscale statistics distinguish visual assessments COVID‐19 classes Machine/deep learning algorithms predict severity from image grayscale attributes Algorithmic class systems can be established using just five grayscale attributes Confusion matrices provide detailed insight to algorithm prediction capabilities