Projet de fin d'étude : Explainable machine learning for covid-19 chest image classification

Etudiant : ELBOUKNIFY ISMAIL

Filière : Master Big Data Analytics & Smart Systems (BDSaS)

Encadrant : Pr. BOUHOUTE AFAF

Annèe : 2022

Résumé : The world is suffering from a global health crisis known as the COVID-19 pandemic. Therefore, there is an urgent need for rapid detection with clear visualization of infection, with the help of which a suspected COVID-19 patient could be saved. for this we proposed to use AI , we made the diagnosis of COVID 19 based on CT-scans . due to the lack of large open source datasets containing CT images of COVID-19 cases, we have used three separate datasets to construct a new dataset contains 1055 Covid-19 cases and 1270 normal cases. We proposed to segment the lungs since they are our region of interest. We used U-net as a segmentation model that achieved 98 % of dice coefficient on the test dataset , we have compared our CNN model with other architecture such as ResNet50 and DenseNet121 applying 5-fold Cross-validation, our model got the best accuracy score 98.40 ± 0.41 % and f1-score 98.23 ± 0.46 %. and to give confidence to the prediction of the model , we applied XIA techniques to understand the reasons behind the predictions ; we applied Grad-cam and integrated Gradient which is based on the calculation of gradient, and LIME based on the perturbation, these methods allows to detect the infections exist in the lungs as reasons behind COVID-19. To compare between the explanations of each method we propose a ground-truth based evaluation method ,This method compares the visualization output with the groundtruth infections, The comparison is performed by measuring the similarity between the two images. the comparison indicates that Grad-cam was the best in our case . we have shown that the prediction of our model based on reasons validated by experts in the field, this amounts to the explanations obtained by the different XAI methods tested in this work .