COVID-CT-Mask-Net: Prediction of COVID-19 from CT Scans Using Regional Features

We present COVID-CT-Mask-Net model that predicts COVID-19 from CT scans. The model works in two stages: first, it detects the instances of ground glass opacity and consolidation in CT scans, then predicts the condition from the ranked bounding box detections. To develop the solution for the three-class problem (COVID, common pneumonia and control), we used the COVIDx-CT dataset derived from the dataset of CT scans collected by China National Center for Bioinformation. We use about $5\%$ of the training split of COVIDx-CT to train the model, and without any complicated data normalization, balancing and regularization, and training only a small fraction of the model9s parameters, we achieve a $\mathbf{90.80\%}$ COVID sensitivity, $\mathbf{91.62\%}$ common pneumonia sensitivity and $\mathbf{92.10\%}$ normal sensitivity, and an overall accuracy of $\textbf{91.66\%}$ on the test data (21182 images), bringing the ratio of test/train data to \textbf{7.06}, which implies a very high capacity of the model to generalize to new data. We also establish an important result, that ranked regional predictions (bounding boxes with scores) in Mask R-CNN can be used to make accurate predictions of the image class. The full source code, models and pretrained weights are available on \url{https://github.com/AlexTS1980/COVID-CT-Mask-Net}. %One of the challenges of training a machine learning model to detect the presence of COVID-related areas in CT scans is the scarcity of segmented data. We present the COVID-CT-Mask-Net, a COVID19 detection model based on instance segmentation of COVID-related areas in CT scans. The model is first trained to segment instances of two types of COVID correlates: ground-glass opacity and consolidation. Then, this model is augmented with three classification modules to predict COVID from the regional features in CT scans. Our model achieves the state-of-the-art accuracy in predicting COVID in patients, it is conceptually simpler, requires a smaller dataset for training compared to other machine learning models, and does not require tricks to balance the data.