Robust Deep Learning Methods for MR(S)I Characterisation of Brain Lesions

The focal point of the PhD is developing novel methodologies for imaging characterisation of brain lesions. Many of the brain-related pathologies are diagnosed and followed-up from magnetic resonance imaging (MRI) scans by assessing the brain lesions. A lot of research has been done in the lesion detection, segmentation, classification, or other forms of lesion assessment. However, due to the scope and diversity of the acquisition protocols, brain pathologies, clinical needs, or specific research objectives, many challenges remain when addressing the robust characterisation of brain lesions in real-world MRIs. The thesis explores different deep learning techniques with the goal of finding the optimal solution for a robust segmentation, classification, and longitudinal analysis of brain lesions in multiple sclerosis (MS).

More recently, magnetic resonance spectroscopic imaging (MRSI) has also become relevant, especially for the assessment of brain tumors. Spectroscopic scans contain valuable metabolic tissue information, but are often noisy and prone to artefacts. One of the main open problems is the qualitative assessment of the acquired data. The work presented in the thesis also includes the deep learning pipeline for automatic and robust quality control of MRSI data. 

Main objectives are addressed in three major points:

O1. Investigate the possibilities for the development of medical imaging software capable of robust detection, segmentation and classification of brain MS lesions; define and develop deep-learning models which can automate and speed up the task.
O2. Analyse the longitudinal patterns of brain MRI biomarkers that reflect MS evolution over time in larger populations; develop tools for longitudinal tracking of the disease.
O3. Research the potential applications and drawbacks of novel MRSI acquisition protocols; propose and develop fast and reliable method for quality control filtering of large numbers of spectra.