JMIR Publications

ABSTRACT

COVID-19 pandemic is becoming one of the biggest unprecedented health crises, and chest X-ray radiography (CXR) plays a vital role in diagnosing COVID-19. Here we present a novel multiple-inputs CNN (MI-CNN) for the classification of COVID CXRs. CXR could be evenly segmented into different numbers (2, 4, and 16) of individual regions. Each region could individually serve as one of the MI-CNN inputs. The CNN features of these MI-CNN inputs would be fused for COVID classification. More importantly, the contributions of each CXR region could be evaluated through the numbers of images that were accurately determined by their corresponding regions in the testing datasets. In both whole-image and LR-ROI (Left and Right-region of interest) datasets, MI-CNNs could get a good efficiency for the COVID classification. Especially, MI-CNNs with more inputs (2-Inputs, 4-Inputs, and 16-Inputs) had better efficiency in recognizing COVID CXRs than the single-input CNN (1-Input). Compared to the whole-image datasets, the efficiency of LR-ROI datasets became about 4% lower in accuracy, sensitivity, specificity, and precision (over 91%). However, by the contributions of each region, we found one of the possible reasons is that some of the non-lung regions (e.g., R16 regions) falsely gave positive contributions to COVID classification. MI-CNN with LR-ROI could provide a more accurate evaluation of the contribution of each region and COVID classification. Additionally, the right-lung regions gave more contributions in the classification of COVID CXRs, while the left lung regions provided more contributions for Normal CXRs. Overall, MI-CNNs could achieve higher accuracy with the increasing number of inputs (e.g., 16-Inputs). It could assist the radiologists in identifying COVID CXRs and screen the critical regions related to the COVID classifications.