Case Study: Covid Detection in RXs

We trained a Residual Network to classify RX images according to their diagnostic outcomes: healthy lungs vs. viral pneumonia vs. bacterial pneumonia. Training datasets are freely available here and here. We opened this model for free use to Health Researchers and Institutions.

X-Ray analysis requires a radiology expert and takes significant time. As a response, here we offer our take on an AI system that does automated analysis of CT scans and automatically classifies VIRAL and BACTERIAL pneumonia presence in real time.

For more information on how this model works, or if you represent a health authority that believes this model might be useful, please write to us at [email protected] We would be happy to share free access to this model to any health authority that requests it.

DISCLAIMER: Please note that this is not a recognized clinical diagnostic procedure, and this tool is not meant to replace medical diagnosis and screening.

60%
or

FAQ

Q: Please can you share the clinical findings and accuracy ratio?

A: Confusion Matrix based on the data set is below:

Confusion between Bacterial and Viral is a bit higher (understandably) than confusion between No Pneumonia and Any Pneumonia.

Q: What is your sample size used for training?

A: Dataset for training:

- 5421 X Ray Scans of Bacterial Pneumonia
- 487 CT Scans of Bacterial Pneumonia
- 4751 X Ray Scans of Viral Pnueomina (includinv Covd19 - 643 cases)
- 352 CT Scans of Viral Pneumonia
- 13243 X Ray Scans of Heatlhy Lungs
- 890 CT Scans of Heathly Lungs

Dataset for validation (also used for the Confusion Matrix above):

- 1191 balanced X Rays and CTScans of the 3 categories that the system has not seen before.
- 0.83 % in classification accuracy in test set.
- 0.98% in classification accuracy where system predicts over 90% Certainity (System reports the certainty of every individual category / prediction).

Q: Do you have clinical protocols established?

A: No. This is up to the health practitioners. We won't produce a clinical protocol.

Q: Why is this model not classifying Covid19 vs non Covid19?

A: In an ideal world, we would have more data to train on, but we don't. We did however had good enough data on Viral Pneumonia vs Bacterial Pneumonia vs Heathly lungs, so decided to take this route.

Covid19 samples would show in the model as Viral Pneuomonia. This does not mean that all viral Pneumonias are Covid19.

Q: How can I help?

A: If you have access to Covid19 positive Xray chestscans, please commit it in the following repository here. The more data is available for teams (including DatumCon) the better the models will be.

Evaluating your Images

You can find below a list of RX Scans. You can zoom on each uploaded image by clicking on it. Processing metadata is reported, together with classification, as assessed by our Residual Neural Networks. DISCLAIMER: Please note that this is not a recognized clinical diagnostic procedure, and this tool is not meant to replace medical diagnosis and screening.

File Type:

Processing Time:

Prediction:

Bacterial Affinity:

Viral Affinity:

Negativity Affinity: