Model Revision Record
Deployment Status: Prototype
Revision: Rev-0
Date: 26-June-2021
Revised by: Woza.Work
Details: First release.
Known Issues:
1- The app can be slow at times, taking around 10 seconds to render a result. The backend code needs to be optimized.
Purpose
This web app uses computer vision to automatically detect covid-19 on chest x-rays. It classifies each image into four classes: negative for pneumonia, typical, atypical and indeterminate. The app also highlights areas of the image where opacity is present.
This is a prototype that has not been field tested. Please use it for educational and demonstration purposes only.
Input
1- Submitted images should be in jpg or png format.
2- The dicom and tiff formats are not supported.
Output
1- The output is a probability score for each of the 4 classes. For example: Typical 0.98 - means that the model is 98% certain that the patient has covid-19.
2- The app also draws a box around areas where opacity is present.
Dataset Summary
The two models that power this app were trained using data from the Kaggle SIIM-FISABIO-RSNA COVID-19 Detection competition.
Source image format: Dicom
Training image format: jpeg
Qty of training images: 999
Num patients: 999
Num women: 9999
Num men: 999
Min patient age: No info
Max patient age: No info
Patient ethnicity: No info
The data is licensed for research purposes only.
Validation Performance
mAP0.5 private leaderboard score: ---
Notes on Algorithmic Bias
1- The dataset did not include patient ages, but looking at the images it appears that there are no chest x-rays of children. Therefore, this model should not be trusted to provide accurate results on child x-ray images.
2- Also, if the dataset images were mainly of senior citizens, who are more susceptible to covid infections, this model could have developed a bias against younger people.
Misc Info
1- The model was trained on jpg images that were created from dicom source images, using a specific conversion process. In practice, users may not use the same dicom to jpg conversion process or they may submit photos of raw x-ray films that were taken with various cameras under various lighting conditions. These real-world factors could impact the performance of the model in unknown ways.
2- The app has been optimized for use on mobile devices.
3- This demo will be live only until the end of September 2021. The code is available on Github. You are welcome to use the code to host this app on your own server.
Documentation
The design code and the step-by-step process used to train and test the models has been published on Kaggle. You can find the open source notebook here. The test results are also available in the notebook.
Contact
Email: contact -at- woza -dot- work
Ref: Covid-19 CXR Analyzer