Diagnostic automatique des AVC par Deep Learning avec Validation des Experts

Abstract

This Master’s thesis presents the development of an intelligent system for automatic stroke detection using brain CT scan images, combining deep learning with an interactive web platform that integrates medical expert participation. The core objective was to design a practical and accurate solution that could be applied in real clinical scenarios. At the center of the system are several convolutional neural network (CNN) models, including ResNet50, DenseNet121, and a custom CNN, all trained on both public and local datasets to classify strokes into three categories: ischemic, hemorrhagic, and normal. Thanks to transfer learning, preprocessing, and early prediction logic, the custom model achieved the best results with a global accuracy of 98.44%, outperforming ResNet50 (96.60%) and DenseNet121 (96.36%). Beyond model performance, the key innovation lies in the creation of a fully integrated web platform built with Django. Users can choose between a free automated analysis or a paid expert review. When expert analysis is selected, a doctor receives the scans, validates or corrects the prediction, provides a diagnosis, and can even communicate directly with the patient through a built-in real-time chat. This feature adds a human layer to the process and strengthens clinical relevance. The project thus offers a hybrid approach that brings together AI automation and expert medical insight, within a secure and usable platform designed for real-world diagnostic support.