Automatic Localization of Bounding Boxes forSubcortical Structures in MR Images UsingRegression Forests

Abstract: Manual delineation of organs at risk in MR images is a very time consuming task for physicians, and to be able to automate the process is therefore highly desirable. This thesis project aims to explore the possibility of using regression forests to nd bounding boxes for general subcortical structures. This is an important preprocessing step for later implementations of full segmentation, to improve the accuracy, and also to reduce the time consumption. An algorithm suggested by Criminisi et al. is implemented and extended to MR images. The extension also includes using a greater pool of used feature types. The obtained results are very good, with an average Jaccard similarity coecient as high as 0.696, and center mean error distance as low as 3.14 mm. The algorithm is very fast, and is able to predict the location of 43 bounding boxes within 14 seconds. These results indicate that regression forests are well suited as the method of choice for preprocessing before a full segmentation.