The model classifies chest CT scans as Covid or Non-Covid images with exceptional sensitivity and accuracy. Trained and tested on very diverse datasets, the network performance generalises very well reliably exceeding accuracies of above 97%. The model delivers a fast and reliable solution for initial screening for Covid-19.
Enough fast and reliable testing capacities are crucial in the current Covid-19 pandemic. While relying on swab sample testing kits can lead to shortenings and relatively long waiting times until a result is delivered, this method simply uses chest CT scans and produces instant test results by analysing if Covid-19 symptoms are present in the patient's lung. The model can be used for training of radiologists and also (after further testing) as a stand-alone solution. Since the development of lung symptoms after initial infection takes a few days, the system is best used on patients who are already symptomatic.
The following is a short manual for the use of the program. A setup manual is also included as a separate .txt file.
The model provides a quick Covid classification analysis of the provided chest CT images after following these steps:
1) Press "Select Folder" and change the directory to the folder that holds the relevant CT scans as .png or .jpg files.
2)Press "Format Images" to convert the images to the target file type and size.
3)Press "Segment Images" to automatically segment the inner lung area from the images. If segmentation results seem suboptimal try changing the brightness value next to the button.
4)Press "Diagnose Images" and the system will predict the Covid factor for each image. It will also show an absolute certainty value and a color matching that value (green = very certain, yellow = certain, red = uncertain).
If you are testing system accuracy on images that are already labeled, you can click one of the checkboxes next to the Diagnose button to test Covid images or Non-Covid images and the system will automatically run through all images and calculate the total prediction accuracy.
Chest CT scans in .png or .jpg format
If requested a DICOM converter can be added later.
The training and testing is based on multiple chest CT scan datasets to add as much diversity as possible. However, due to the novelty of the disease, the publicly available data is limited and a total amount of roughly 900 Covid and 900 Non-Covid images was used for training and testing. Nevertheless, the performance seems to hold up very well on independent test sets.
Model performance metrics
The performance metrics on a very diverse testing dataset from were: Accuracy 0.970443 F1-Score 0.969697 AUC -Score 0.970748 These metrics exceed other competitors and seem to translate very well to other datasets; it should be worth testing on your data or even in the wild.