Covid19action-radiology-CXR

The outbreak of COVID-19 in Wuhan, China in December 2019 has rapidly spread across other countries in the world and has been declared as a global pandemic by WHO on 11th March, 2020. COVID-19 continues to have adverse effects on the health and economy of the global population and has brought immense pressure on the health care systems of the developing as well as  developed countries. A critical step in the fight against COVID-19 is the early detection of the patients which not only aids in providing quicker treatment but is also useful in stopping the spread of the virus through proper isolation of the patient.

Although the reverse transcription polymerase chain reaction rt-(PCR) test of the sputum is considered as the gold standard for COVID-19 diagnosis, it is a time-consuming process and, in some cases, has led to high False Negatives. The relatively low-cost, wider availability and faster methods for sanitization of the equipment makes Chest X-ray imaging (CXR) a promising modality which is commonly being employed for treatment planning, tracking the progression of the disease and could also be useful in the detection of COVID-19.

Our goal is to create a single repository by collating various publicly available CXR image sources and provide a standardized split for performing a five-fold cross-validation based training and a separate held out test set for evaluating different AI models. We hope that it will provide a unified platform for researchers to perform a fair comparison of different AI models. This dataset comprises images acquired from different geographical regions using different scanners and at varying resolutions.

The primary task is to classify a given CXR image into “COVID-19”, “Other Pneumonia” and “Non-pneumonia” classes. The “Non-pneumonia” class contains images from healthy subjects as well as subjects suffering from diseases other than pneumonia. Additionally, the ground truth (GT) annotations of 13 radiological observations are also available for the images for a subset of images. These may be used to define an additional auxiliary task for the model or pre-training it to improve generalization.