Prédiction de l’efficacité de la chimiothérapie appliquée au cancer du sein par le traitement d’images et le Deep Learning
Mohammed El Adoui · HAL (Le Centre pour la Communication Scientifique Directe) · 2021
Breast cancer is one of the most common diseases in women around the world. This cancer is the leading reason for death in women aged 35 to 70 years old. The growth of cases of breast cancer, as well as a large number of imaging examinations carried out in recent years, provided the development and made it possible to automate several medical imaging techniques. Magnetic resonance imaging (MRI) exams present a great interest to radiologists. Indeed, MRI performs to have a temporal follow-up of the breast tumor thanks to the multiple information and sub-modalities produced by this robust medical imaging modality.In this thesis work, the primary purpose is to help oncologists and radiologists to predict the breast tumor response to chemotherapy. Technically, this could be made by comparing MRI scans before and after the 1st chemotherapy. Such predictions will help to make quick decisions, based on how a breast tumor responds to chemotherapy from the start of therapy.We conducted in-depth research in the literature related to classical imaging approaches, which led us to propose and implement a first technique called Parametric Response Map (PRM). This method is coming off on two primary steps: segmentation and three-dimensional registration of images acquired before and after the 1st chemotherapy. PRM allowed a voxel-by-voxel comparison of the tumor volume. This method produced an easy-to-read color map identifying intra-tumoral regions that responded to treatment (positive response), unresponsive regions (stable), and regions that experienced aggressiveness progression (negative response) indicating the percentage of each tumor's zone.Then, we used deep neural networks for segmenting tumor’s volumes and predicting their response to chemotherapy in an automatic way based on several databases provided by several international institutionsThe promising results of this study show an accuracy value of 89% using the PRM method, and an average accuracy of 93% using Deep Learning across multiple datasets. To our knowledge, these results put themselves above all the presented results in the literature. The standard reference used to validate all the proposed methods is the pathological complete response (pCR) obtained for each patient included in this study.