Automatic Stroke Lesion Segmentation With Limited Modalities

MRI Brain image segmentation is a valuable tool in the diagnosis and treatment of many different types of brain damage. There is a strong push for development of computerized segmentation algorithms that can automate this process because segmentation by hand requires a great deal of effort by a highly skilled professional. While hand segmentation is currently considered the gold standard, it is not without flaws; for example, segmentation by two different people can provide slightly different results, and segmentation by hand is labor intensive. Due to these flaws, It is desirable to make this process more consistent and more efficient through computer automation. This project investigates four promising approaches for the automatic segmentation of brain MRIs containing stroke lesions found in recent literature. Two of these algorithms are designed to use multiple modalities of the same patient during segmentation, while the other two are designed to handle one specific modality. The robustness of each to limited, or different, image sequences than they were originally designed for will be tested by applying each to two datasets that contain 24 and 36 patients with chronic stroke lesions. These tests concluded that performance for the multi modal algorithms does tend to decrease as input modalities are removed, however it also revealed that FLAIR imaging in particular seems to be especially valuable for segmenting stroke lesions. In both multi-modal algorithms while there was an overall drop in Dice scores, any testing that included FLAIR images performed significantly better than any other tests. The single channel algorithms had difficulty segmenting any modalities different from the specific one that they were designed for, and were generally unable to detect very small lesions.