ABSTRACT
Background The COVID-19 pandemic is becoming one of the largest, unprecedented health crises, and chest X-ray radiography (CXR) plays a vital role in diagnosing COVID-19. However, extracting and finding useful image features from CXRs demand a heavy workload for radiologists. Objective The aim of this study was to design a novel multiple-inputs (MI) convolutional neural network (CNN) for the classification of COVID-19 and extraction of critical regions from CXRs. We also investigated the effect of the number of inputs on the performance of our new MI-CNN model. Methods A total of 6205 CXR images (including 3021 COVID-19 CXRs and 3184 normal CXRs) were used to test our MI-CNN models. CXRs 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 then be fused for COVID-19 classification. More importantly, the contributions of each CXR region could be evaluated through assessing the number of images that were accurately classified by their corresponding regions in the testing data sets. Results In both the whole-image and left- and right-lung region of interest (LR-ROI) data sets, MI-CNNs demonstrated good efficiency for COVID-19 classification. In particular, MI-CNNs with more inputs (2-, 4-, and 16-input MI-CNNs) had better efficiency in recognizing COVID-19 CXRs than the 1-input CNN. Compared to the whole-image data sets, the efficiency of LR-ROI data sets showed approximately 4% lower accuracy, sensitivity, specificity, and precision (over 91%). In considering the contributions of each region, one of the possible reasons for this reduced performance was that nonlung regions (eg, region 16) provided false-positive contributions to COVID-19 classification. The MI-CNN with the LR-ROI data set could provide a more accurate evaluation of the contribution of each region and COVID-19 classification. Additionally, the right-lung regions had higher contributions to the classification of COVID-19 CXRs, whereas the left-lung regions had higher contributions to identifying normal CXRs. Conclusions Overall, MI-CNNs could achieve higher accuracy with an increasing number of inputs (eg, 16-input MI-CNN). This approach could assist radiologists in identifying COVID-19 CXRs and in screening the critical regions related to COVID-19 classifications.
ABSTRACT
Background: Chest CT had high sensitivity in diagnosing novel coronavirus pneumonia (NCP) at early stage, giving it an advantage over nucleic acid detection in time of crisis. Deep learning was reported to discover intricate structures from clinical images and achieve expert-level performance in medical image analysis. To develop and validate an integrated deep learning framework on chest CT images for auto-detection of NCP, particularly focusing on differentiating NCP from influenza pneumonia (IP). Methods: 35 confirmed NCP cases were consecutively enrolled as training set from 1138 suspected patients in three NCP designated hospitals together with 361 confirmed viral pneumonia patients from center one including 156 IP patients, from May, 2015 to February, 2020. The external validation set enrolled 57 NCP patients and 50 IP patients from eight centers. Results: 96.6% of NCP lesions were larger than 1 cm and 76.8% were with intensity below -500 Hu, indicating less consolidation than IP lesions which had nodules ranging 5-10 mm. The classification schemes accurately distinguished NCP and IP lesions with area under the receiver operating characteristic curve (AUC) above 0.93. The Trinary scheme was more device-independent and consistent with specialists than the Plain scheme, which achieved a F1 score of 0.847, higher than the Plain scheme (0.774), specialists (0.785) and residents (0.644). Conclusions: Our study potentially provides an accurate early diagnosis tool on chest CT for NCP with high transferability, and shows high efficiency in differentiating NCP and IP, helping to reduce misdiagnosis and contain the pandemic transmission.