Medical imaging has gained unprecedented momentum in recent years as it offers non-invasive investigation of several organs for the purposes of diagnosis, monitoring and even prognosis. It is positioned as a central tool in numerous large-scale studies involving several inclusion centers. These multicenter studies, however, generate heterogeneous datasets because: a) the images come from different manufacturers, b) they are acquired with sometimes variable settings, c) they come from centers with different degrees of expertise and d) they are acquired on patients who are more or less difficult to scan (movements linked to a pathology, inability to hold an apnea, etc.). This great variability requires good standardization of acquisition but also of data processing in order to provide quality biomarkers to clinical investigation centers. This analysis includes the step of extracting diagnostic or prognostic biomarkers but also preliminary steps which are nevertheless of major importance: a) the pseudonymization of the images in order to be able to transfer them to the analysis center in agreement with the legislation in force and b) the quality control of the images which must accompany the quantitative biomarkers in order to guarantee their relevance.

In this context, the objective of this thesis is to propose a brain MRI image analysis pipeline based on machine learning and which includes:

  1. A step of anonymization of DICOM metadata and tags displayed on images
  2. A quality control (QC) step which will be carried out as a preamble to quantitative analyzes and expert radiological readings to allow pre-sorting and save time for clinicians