The model is a Pre-screening Opacity Detector for COVID-19 which is trained on CXR images from various sources such as the NIH, RSNA, Vuno Datasets

Value Proposition

Diagnosis of Opacities is essential for the first stage of screening for COVID-19. The ease of availability of X-Rays throughout the world makes the model more suitable for real life applications. The model helps reduce direct contact of radiologists with patients and solves the problem of limited access to testing kits.

Intended Use

The model takes a single or multiple DICOM files as input. The intended usage is for diagnosis of lung opacities for early diagnosis of COVID-19. The model is useful for pre-screening of patients with common symptoms of the virus like fever, cough, breathlessness etc. The patients can then be tested using more reliable methods like PCR, thereby speeding the procedure and avoiding shortage of testing kits.

Limitations

1. Limited access to PA/AP Chest X-rays of COVID patients for training could lead to poorly generalized results in some cases.
2. The model is yet to undergo rigorous testing and hence concrete and accurate results cannot be guaranteed always

Information on training data

The model uses PA and AP CXR images from multiple sources including some famous Pneumonia datasets like NIH, RSNA, etc. The model was specifically trained on COVID-19 cases from various hospitals across the US, China and Italy provided by Joseph Paul Cohen (https://github.com/ieee8023/covid-chestxray-dataset). A separate dataset of Spanish COVID-19 Patients was added later to cover up for shortage of data.

Model performance metrics

1\. Accuracy- 0.9561
2\. Precision- 0.9538
3\. Recall- 0.9431
4\. F1-Score - 0.9507
5\. AUC - 0.9841

Performance Curves

TrainingGraph

Model History

Model available on Arterys.

Darshan Deshpande

Darshan Deshpande

Learn more