Magnetic resonance imaging is now an essential tool for studying the brain and diagnosing many neurological diseases. It allows the brain's structure and abnormalities to be observed non-invasively. Among the various acquisition sequences, the FLAIR sequence plays an essential role in detecting brain lesions. These images are widely used in clinical practice, particularly for conditions such as multiple sclerosis, brain tumors, and strokes. However, their automatic analysis remains difficult due to the low contrast between tissues, the limited signal-to-noise ratio, and the variability between imaging devices. In neuroimaging, quantitative analysis of the brain increasingly relies on automatic segmentation tools capable of identifying anatomical structures and pathological regions.
Most of these methods were designed for T1-weighted images, which are considered the gold standard for anatomical segmentation. However, this sequence is not systematically acquired in clinical protocols, which limits the use of existing tools. In many cases, the FLAIR sequence is the only one available, but conventional approaches, designed for T1w, fail to provide reliable results. This thesis addresses this issue and explores the possibility of performing fine segmentation of the brain using FLAIR images alone. The objective is to develop robust methods capable of extracting relevant anatomical and pathological information from this sequence, while adapting to acquisition variability and the presence of lesions. The work carried out has led to the development of tools specifically adapted to FLAIR, designed to improve the performance of preprocessing and segmentation steps in a realistic clinical setting.
The results obtained show that it is possible to obtain fine and consistent brain segmentations from FLAIR alone, paving the way for the exploitation of vast clinical databases that have been underutilized until now. These approaches also demonstrate the value of designing specialized models for each modality rather than generic methods that are independent of the sequence. Ultimately, the tools developed could be integrated into the volBrain online platform (https://volbrain.net) to facilitate their dissemination to the scientific and clinical community.