This work is for didactic purposes (train & classify neural network) and should not be used for real diagnosis
Classify COVID-19 CT images using ResNeXt and other networks
I have used the data given at https://github.com/UCSD-AI4H/COVID-CT/tree/master/Images-processed to train a neural network. All images go into one single directoy (./images). The truth information is stored in the csv files for training and validation
After 300 Epochs, the accuracy was above 98% with a confidence of > 95%.
Accuracy was calculated by counting the True Positve as well as the True Negative classifications and divide the sum by the total number of images calssified.
Confidence is calculated by averaging the network prediction for each classification
The code is based on ResNeXt50_32x4d with an input image size of 512x512 On a NVIDIA GTX1660, one epoch running through the ~700 images takes about 1 minute, the 300 epochs took about 32 hours
Loss Evolution:
Covid-19 fidning (True Positive):
No finding (True negative):
