A Pipeline For Segmenting and Classifying Brain Lesions Caused By Stroke: A Machine Learning Approach

Brain injuries caused by strokes are one of the leading causes of disability in the world. Current procedures require a specialized physician to carefully analyze an MRI image before making a diagnosis, but the procedure can be costly and time-consuming. Artificial intelligence techniques are playing an important role in the analysis of this kind of images. This work proposes an end-to-end approach in three stages: Pre-processing, for normalizing the images to the standard MNI space, as well as inhomogeneities and bias corrections; then lesion segmentation using a CNN network trained for Cerebrovascular Accidents (CVA) and feature extraction; and, at the end classification for determining the vascular territory within which the lesion occurred. Four Deep Neural Network (DL) architectures, as well as four Shallow Machine Learning (ML) models were designed, analyzed and compared based on their performance scores, reaching an accuracy of 84% with the DL models, and 95% with the Shallow ML models. The proposed method can be used as an automatic predictor for decision-making to assist health service providers of stroke lesion identification from MRI images.