DeepCOVID-XR: An Artificial Intelligence Algorithm to Detect COVID-19 on Chest Radiographs Trained and Tested on a Large US Clinical Dataset

DeepCOVID-XR is a model for COVID-19 detection using chest x-ray images. It is an ensemble of convoutional neural network (CNN) models including DenseNet-121, EfficientNet, Inception, InceptionResNet, ResNet-50 and Xception.  Each model was pretrained on the NIH dataset. The COVID-19 chest x-ray image dataset we used consists of more than 14,000 images of which about 4000 are COVID-19 positive images. The ground truth is obtained with the real-time polymerase chain reaction test. Each image is also preprocessed with U-Net to crop out the lung section from the original image to enhance the performance. Each CNN model is then fine tuned with the preprocessed dataset. To evaluate  our model,  300 random images from the test set are also evalueated by five experienced thoraic radiologists.

The GRAD-CAM of our model shows that the Our ensembled model can reach an accuracy of 90% and an AUC of 0.9 when tested. When compared to the radiologists’ reading, our model has a slightly better accuracy of 82% and AUC of 0.88. The accuracy of the radiologitsts’ reading ranges from 76% to 81%. The consensus accuracy is 81% and the AUC is 0.85.

The results of our model demonstrates potential benefit for healthcare system by reducing exposure to virus and automatically recognize suspicious patients for further testing. For this purpose, intitutions around the world can further fine tune and test the model with other datasets.

The paper for this project has been accepted by Radiology and published online.

Check out our github repository for DeepCOVID-XR.