Using Deep Learning for MRI Stroke Lesion Segmentation

Abstract

The project aims to exploit Deep Learning techniques for the segmentation of MRI images to assist radiology specialists in the detection of strokes. The main aim is to improve the detection process using the advantages of artificial intelligence. Several approaches were explored, including the creation of a customized CNN model, the use of learning transfer and the ensemble learning. An in-depth comparative study was carried out to evaluate the performance of the different models obtained, focusing in particular on Dice, IoU and Precision scores. Ultimately, after careful evaluation, two of our models, Efficientnetb3-50 and Model HT-FLAIR, were selected and proved to be the best choice thanks to their better scores.