Dépistage et classification des tumeurs pancréatiques par apprentissage automatique et profond.

Abstract

The aim of this project project is to link artificial intelligence technologies Machine Learning and Deep Learning to the field of medicine and specifically to the screening of pancreatic cancer. The study was conducted on the basis of (02) two distinct approaches which are screening by medical analysis and screening by medical imaging. The first approach consists in making predictions on the basis of urine tests performed on patients and linking its data to learning models optimized by two methods (Data Augmentation and Cross Validation with the K-fold technique) so that they can make accurate and reliable predictions on other external data to be provided. In a comparative study between the two algorithms used (Random Forest Classifier ) and (K-Nearest Neighbors), it was found that the Random Forest Classifier model provided the best results, i.e. 98.87 % in the training phase and 99.33 % in the test phase. Regarding the second approach, we used the technique of Transfer-Learning which proposes models that are already trained on the Imagenet database such as : Xception and Densenet121 which are known to provide high accuracies on a large number of classes. Comparing the two models used, we selected the Xcpetion model for the high performances it provided i.e. : 98.14 % - (Accuracy) and 3.92 % - (Loss). By adapting its models to our dataset in the form of CT-SCAN, they provided very satisfactory results that we wanted to project on a platform that offers interesting tools that improve the handling aspect of CT-SCAN and offer a better diagnostic experience.