A stroke is a life-threatening condition when part of the brain does not have enough blood flow due to an artery blockage or brain bleeding. The World Health Organization, known as the WHO, reports that stroke ranked as the second most common cause of death worldwide in 2019, accounting for almost 11% of all fatalities. An extensively utilized method for diagnosing strokes is magnetic resonance imaging (MRI), which offers detailed information on the interior anatomy of the brain. Radiologists manually segment the stroke lesions by analysing the MRI images. However, manual segmentation is a complex and more time-consuming task due to multiple MRI sequences, slices of MRI images, and the brain’s anatomical structure. Therefore, it becomes crucial to develop automatic segmentation methods to help radiologists effectively detect the severity and location of the brain stroke more accurately and timely. Even with significant progress in medical imaging research, automatic stroke lesion segmentation is still tricky. Based on a literature review, this review paper highlights the challenges in semantic segmentation in MRI images. Such challenges include class imbalance based on tiny lesions, fuzzy lesion boundary segmenting, and intraclass and interclass inconsistencies between lesions and healthy tissues. Variations in lesion size and position within MRI slices further complicate semantic segmentation. This review study outlines the current state of research regarding semantic segmentation in MRI images and guides overcoming these obstacles in the future based on ATLAS and ISLES stroke datasets.