Applied AI in Biomedicine Project

Chest X-Rays Image Classification project from the course Applied AI in Biomedicine at Politecnico di Milano

Final grade: 30/30

Data

The provided dataset contais 15470 labeled images of healthy individuals and individuals affected by either Pneumonia or Tuberculosis.

The dataset is unbalanced with the normal, pneumonia and tuberculosis classes representing respectively the 60%, 27% and 13% of the data

We trained and tested a multitude of models, comparing them with the F1 score metric on the validation set

Model	F1 - Normal	F1 - Pneumonia	F1 - Tuberculosis
SVM (HOGs)	0.927	0.953	0.771
CNN Scratch	0.969	0.980	0.882
EfficientNetB2	0.982	0.981	0.936
EfficientNetB3	0.978	0.982	0.921
DenseNet121	0.968	0.976	0.883
VGG16	0.974	0.983	0.900

We chose EfficientNetB2 as our best model and we evaluated it on the test set

Model	F1 - Normal	F1 - Pneumonia	F1 - Tuberculosis
EfficientNetB2	0.975	0.977	0.921

We interpreted the results of our model with different explainability techniques such as Grad-CAM, Occlusion analysis and LIME

This project is licensed under the MIT License.