Semi-Supervised Novel Radiomics Approach for COVID-19 Severity Detection based on Atmospheric Veil Correction

Computed tomography (CT) of the lungs is a widely used imaging technique for detecting lung abnormalities. It plays a crucial role in monitoring the progression of diseases, efficiently allocating scarce medical resources, and strategizing treatment in clinical settings. However, the manual screening of CT images is laborious and prone to human errors. Thus, deep learning (DL) can serve as a supplementary decision-making tool for radiologists. While previous studies have achieved notable success in classifying Coronavirus Disease 2019 (COVID-19), accurately identifying the location of the infection and measuring its severity remains challenging. This study presents a three-step framework for the semi-supervised radiomics classification of COVID-19, identification of the infection region, and automatic assessment of disease severity. The initial step involves identifying the illuminated infected area in the CT image using the intensity dark channel prior (IDCP) technique. Additionally, the number of pixels in the bright region is estimated through atmospheric veil correction (A VC) for the labeled CT images. In the second step, a random forest classifier is trained to predict the classes of the unlabeled CT images, separated based on the number of bright pixels and texture features. The third step involves performing classification using a residual neural network with 50 layers (Resnet-50) and verifying model accuracy with labeled data. The experimental results demonstrate that the proposed method achieves verification results of 98.99%, 97.73%, 100
^%
,98.21 %, and 98.85% in the metrics of accuracy, precision, sensitivity, specificity, and F1 Score, respectively. Additionally, our model's performance surpasses that of current state-of-the-art studies conducted on the same dataset.