Purpose: COVID-19 from SARS-CoV-2 targets the lungs, and thus, there is a pressing need for effective surveillance using chest radiographs (CXR). Our study aims to use a deep-learning based lung field segmentation of CXR exams to aid in the task of COVID-19 diagnosis.
Methods: In the first stage, quadrilateral masks were generated from manually delineated anatomic-point-based regions within the CXR lung. Using an initial cohort of 631 COVID-19 positive patients, a U-Net was trained from the CXR images and masks, split into 64%/16%/20% for training/validation/testing to identify the lung field. In the second stage, the trained U-Net model was used to segment the lung region on CXRs from an additional cohort of 9,126 patients tested with RT-PCR. This dataset was split for training/validation/testing to train and evaluate a DenseNet121 in the task of differentiating COVID-19 positive and negative patients. Classification performance was evaluated using receiver operating characteristic (ROC) analysis with area under the ROC curve (AUC) as the figure of merit. Gradient-weighted class activation mappings (Grad-CAM) were generated for explainability.
Results: On the test set, U-Net segmentation yielded a Dice coefficient of 0.96 (95% CI: 0.90, 0.98) and average Hausdorff surface distance of 14.3 mm (11.7 mm, 15.9 mm). On the second test set, the DenseNet in the positive/negative classification yielded AUC values of 0.77 (0.74, 0.80) and 0.76 (0.72, 0.76) with and without the use of the thoracic mask, respectively. While the two curves failed to show a statistically significant difference, ΔAUC of 0.01 (-0.01, 0.03) with P=0.16, the Grad-CAM heatmaps showed reduced influence from outside the lung region when masks were applied to the images.
Conclusion: Automatic lung field segmentation achieved a promising classification performance while limiting predictions from regions outside the lung. Demonstrated performance of the U-Net segmentation indicates its high potential for future COVID-19 diagnosis classifications.
Funding Support, Disclosures, and Conflict of Interest: This research was supported by NIH T32 EB002103, NIH NIBIB contract 75N92020D00021, and the C3.AI Digital Transformation Institute.