In this project I trained a CNN for the classication of Chest X-Ray Images of healty/pneumonia cases, choosing the best one between 3 different approaches:
- a custom (deep) CNN;
- a shallow CNN (SCNNB);
- a pre-trained ResNet50.
I then chose the best model based on various metrics, optimal for medical analysis, and I interpreted the results using gradients-based techniques by using the tf-explain library.
Use this link if you have problems visualizing the notebook.
The dataset used for the project is taken from the Chest X-Ray Image (Pneumonia) kaggle competition.
It consists of 5863 X-Ray images (JPEG format) and 2 categories (Pneumonia/Normal), divided into:
- train: 1341 Normal, 3875 Pneumonia
- test: 234 Normal, 390 Pneumonia
- main.ipynb: notebook containing all the code for the completion of the project.