Image processing techniques for quantification and assessment of brain MRI

Abstract

Magnetic resonance imaging (MRI) is a widely used technique to acquire digital images of the human brain. A variety of acquisition protocols is available to generate images in vivo and noninvasively, giving great opportunities to study the anatomy and physiology of the human brain. In my thesis, image processing techniques for quantification and assessment of these images are discussed. These techniques are applied in the context of brain anatomy and pathology, in particular small vessel disease. Advances in MR imaging have led to a tremendous increase in the amount of data and level of detail that can be acquired. This causes manual assessment of images to become increasingly difficult and time-consuming, thereby threatening the quality of such assessments. Image processing techniques are indispensable to human observers in achieving the best possible results. With the use of 7T MRI, cerebral microbleeds can be visualized on gradient echo images. Owing to their small size, manual detection is time-consuming, rater-dependent, and has a limited robustness and reproducibility. In my thesis, a new method to determine the quality of manual microbleed detection is presented, which is suited for images acquired at 7T and for patients that have multiple microbleeds. Since manual detection is difficult, semi-automated detection is likely to improve the quality of ratings. In my thesis, a semi-automated detection technique for microbleeds is presented. This technique detects potential microbleed locations and its findings are presented to a human observer for the final identification of true microbleeds. By using this technique, the sensitivity and quality of microbleed detection increased and the required human observer time was reduced. The most important finding was the detection of extra microbleeds that were initially missed by human observers, but were confirmed as true microbleeds. This demonstrates the high difficulty involved with manual microbleed detection and stresses the importance of robust and reliable (semi-)automated techniques. A similar semi-automated approach was applied to the detection of cortical cerebral microinfarcts, also known as the “invisible lesion”. Microinfarcts receive high interest, because of their relation with cerebrovascular disease and dementia. Recently, microinfarcts were visualized with 7 T MRI. Manual detection of microinfarcts required approximately 30 to 60 min per patient. A semi-automatic detection technique is presented in this thesis that reduced the required rating time to 5-20 min. Next to this, extra microinfarcts were detected that were initially missed by the human observer. Fully automated image processing techniques were applied to extract the midsagittal plane from brain MR images. The midsagittal plane separates the two hemispheres and is used for many analyses that compare one hemisphere with the other, as well as a preprocessing step for further image processing. Multiple techniques to extract the midsagittal plane were evaluated and compared to manual delineations. In addition, a technique to extract the (curved) midsagittal surface is presented. Since the interhemispheric fissure that separates the two hemispheres is not exactly planar, a plane is not a correct separation. The midsagittal surface captures the natural shape of the interhemispheric fissure and provides a better separation.