Research  >  Research  >  Labs & groups  >  Cardiovascular MRI  >  Introduction to MR physics  >  Analysis


MR physics: Image analysis

The nature of the data produced with MRI is, in many ways, well-suited to subsequent image analysis. Specifically: (i) many different organs can be distinguished with the flexible soft tissue contrast; (ii) resolution can be tailored to specific applications and noise is generally well behaved; (iii) signal can be resolved in the three spatial dimensions to sub-millimeter levels; (iv) one can resolve temporal signal changes at the sub-second level; and (v) one can obtain multiple measurements of the same volume element with different contrast characteristics. The vast volume of data alone often demands computer image analysis for interpretation and display.

Areas of research in image analysis include methods for image segmentation [35,36,37] and registration [37,38]. Image segmentation facilitates (i) volume estimation for characterizing disease (ejection fraction in the heart, tumor volumes), (ii) isolation of a tissue of interest from a three-dimensional data set, notably for the display of vascular anatomy, and (iii) multiple image registration through the identification of common landmarks in the images. Registration facilitates (i) temporal signal analysis of a specified volume within the same study, (ii) comparative studies of the same patient over multiple studies perhaps with data of different contrast weightings or even data from different imaging modalities, and (iii) comparative studies of anatomy across multiple individuals.

The task of image analysis is simplified greatly when attention is paid to the nature of the basis images. As noted in sec. iii-C, image sequence parameters can be selected to maximize contrast between tissues, facilitating segmentation. Acquisition can be synchronized to motion or priority can be given to imaging speed to reduce the demands on registration algorithms. Similarly, registration of multiple images with different contrasts is aided by acquiring lines of k-space from the different images in a time-interleaved manner. For segmentation, image resolution should be selected to minimize the difficulties associated with partial voluming, where multiple tissues are present in the same voxel, based on the sizes of the structures of interest. Finally, the additive, Gaussian nature of the noise can be exploited to optimize classification strategies in segmentation; as a caveat, the noise takes on a Rician distribution in magnitude images most often used for segmentation as a result of the magnitude operation.