COVID-19 CXR Analyzer
Model Revision Record
Model Name: exp145-best.pt
Deployment Status: Prototype
Revision: Rev-P0
Date: 18-July-2021
Revised by: vbookshelf
Details: Released for demonstration.
Known Issues:
1- The latency (prediction time) can sometimes be more than 5 seconds. This needs to be reduced to less than 3 seconds.
Purpose
This web app uses computer vision to detect and localize COVID-19 on chest x-rays. It classifies each image into one of two classes:
- Negative for pneumonia
- Typical for COVID-19
Input
1- Submitted images should be in jpg or png format.
2- The dicom and tiff formats are not supported.
Output
The app outputs one of two predicted classes - Negative for pneumonia or Typical for COVID-19. When the app predicts 'Typical for COVID-19', it also draws bounding boxes around areas where it detected opacity.
Dataset Summary
I fine tuned a Yolov5 model using data from the Kaggle SIIM-FISABIO-RSNA COVID-19 Detection competition.
The original dataset has four classes:
- negative for pneumonia
- typical, atypical and indeterminate for COVID-19
Source image format: Dicom
Training image format: jpeg
Total images: 6334
Num patients: 3261
Num female patients: 1624
Num male patients: 1637
Num female images: 2770
Num male images: 3564
Min patient age: No info
Max patient age: No info
Patient ethnicity: No info
You can find the dataset license info here.
Validation Performance
Confusion Matrix
Classification Report
I used image augmentation during training. This helps the model to generalize. Also, if images of the same patient did appear in both the train and validation sets, augmentation reduces the similarity between those images.
Misc Info
1- The dataset did not include patient ages but it appears that there are no chest x-rays of children. Therefore, this model should not be trusted to provide reliable results on child x-ray images.
2- In practice users could submit images of varying quality. These images could have been taken using different types of x-ray machines or even be photos of x-ray films taken with cellphone cameras. These and other real-world factors could reduce the accuracy of the model.
3- The app has been optimized for use on mobile devices.
4- This demo will be live until 31 August 2021. The code is open source. You are welcome to use it to host this app on your own server.
Documentation
1- The frontend and backend code is available on GitHub.
2- My code is available under an MIT License. But please note that the model can't be used commercially because I trained it using competition data that is licensed for research use only.
Contact
Email: contact -at- woza -dot- work
Ref: COVID-19 CXR Analyzer