Evaluation of deep learning transformers models for brain stroke lesions automatic segmentation

Brain stroke represents a leading cause in long-term disability worldwide, stroke rehabilitation research is focused on the understating of the relationship between brain, behavior, and recovery, using as a basis brain changes generated after a stroke, this allows for precise diagnostic and possible predictions in terms of functional outcomes. Neuroimaging represents the main resource for brain stroke research and therapies, it is of particular interest high-resolution T1- weighted (T1w) anatomical MRIs, which are used to evaluate/examine structural brain changes after stroke episodes. Several techniques have been developed in order to accurately calculate or approximate the percentage between lesions and critical brain structures, this step constitutes a paramount step for precise lesion annotation. Despite the technological progress or advance, to date, manual lesion tracing by a team of experts in neuroimaging remains as the gold standard to draw valid clinical inferences for lesion segmentation. The following work proposes a review of the machine and deep learning models that have been developed focusing in the transformers algorithm which is a state of the art method based on the self attention mechanism that has outperformed recurrent neural networks in terms of evaluation metrics such as the dice value, being able to capture long distant dependencies which is a fundamental step when processing 3D volumes, formed by a stacked 2D MRI images. The models were tested using the ATLAS dataset (Anatomical tracing of lesions after stroke) which is an open source data set of T1-weighted MRIs with manual segmented brain lesions.