Stroke lesion segmentation in MRI datasets using deep learning

Strokes, blood-flow related brain damage, are currently the third cause of death worldwide. After a stroke incident, the clinicians asses the damage caused to the brain using medical imaging techniques, such as magnetic resonance imaging (MRI). While visual interpretation is still the most common and accepted way to analyse brain images, they are also subject to computer analysis such as delimiting regions of interest: in the case of strokes, segmenting the lesion area. Currently, this segmentation task is carried on by hand, pixel by pixel, slice by slice, using drawing tools. This thesis has for objective to help the clinical staff with this task using automatic models from the field of artificial intelligence: deep learning models, and more specifically fully convolutionnal neural networks such as 3D U-Net variants.