Severity of COVID-19: A Multimodal Guided Evaluation with Classification and Segmentation

Everyone across the globe has witnessed the effects of the COVID-19 pandemic. While numerous models have been developed to identify diseases, only a limited number have progressed to the stage of evaluating infections. The objective of this study is to assess the extent of infection in a given region by assigning a severity score following the classification of CT scan images. The model comprises of two iterations pass 1 and pass 2. The aim is to design a model with two passes for the severity evaluation. Where, pass 1 is the categorization process and pass 2 for the area-based severity evaluation phase. In the first pass, a pre-trained Mobile Net model is employed in conjunction with a Support Vector Classifier (SVC). In the second pass, watershed-based segmentation techniques are utilized, and area-based feature extraction is performed on the outcome of the second run for assessment purposes. In the training and testing phase of pass 1, a combined total of 349 COVID and 397 Non-COVID photos were utilized, following an 80:20 ratio. Additionally, pass 2 involved the validation of severity evaluation using 34 COVID-positive images, which yielded an AUC of 0.93. The conclusion of the study proposes a severity classification system, which includes three categories: extremely severe (above $5 \%$), severe ($2 \%-5 \%$), and moderate (up to $2 \%$). The combined model, incorporating both passes 1 and 2, achieved an accuracy of $91 \%$, precision of $100 \%$, and recall of $87 \%$. The task will provide further assistance to the pulmonologist.