Covid-non Covid detection, with two experiments: RES_NET transfer learning vs custom CNN trained from scratch.
The aim is to identify Covid from x-ray photos of lungs, training the network on a non-covid covid dataset (augmented with techniques such as crop, translation, contrast jitter ...).
Completed
-
Flavio Lorenzi
-
Nicolò Mantovani
-
Michele Ciciolla
Idea, part of dataset and of the transfer learning experiment is taken from here: (shervinmin/DeepCovid)
Dataset, still too small, it has been increased thanks to some Data Augmentation techniques described in the document (pdf).
COVID
NO COVID
Very good results for each experiment, with an inference accuracy bigger than 0.9 : this is the consequence of a careful tuning of the parameters, but above all, this is due to the fact that it is a Binary Classification (easier to train than a multiclass one).
Custom CNN training-test plot:
Confusion matrix of the Transfer Learning inference experiment, given a blind balanced dataset of 100 imgs:
See other results, indices and successes in the pdf document.
The two final accuracies differ very little (0.95 v 0.93) and therefore it is difficult to judge which model is really better; but we came to the conclusion that it is better to choose transfer learning, because not only does it have better accuracy, but above all it did not require a large computational calculation, performed in a short time and locally, while the second experiment required Colab's gpu to speed up the timing.
In the future, having more images available, it would be interesting to pull the custom network on a larger dataset, in order to seek even better performance; or the network could be extended for a classification that is no longer binary but multiclass, introducing for example the recognition of various pathologies (pneumonia, water in the lungs, fractures, etc.). It would be an excellent pathology detection.