Semi-automatic segmentation of anatomical structures from CT images for radiation treatment planning
Michael D. Grossberg, PhD and Gikas Mageras, PhD
Segmentation of anatomical structures in volumetric images is an essential component of radiation treatment planning, yet is increasingly time consuming with the growing number of images from modern equipment. Further, radiotherapy accuracy is limited by the variability of manual segmentation. We propose the development and validation of a method of semi-automatic image segmentation. The long-term goal is to improve medical experts' productivity and consistency in delineating normal radiation-dose-limiting tissues. The specific aims are to 1) develop a method and software tool for segmentation of 3D stacks of computed tomography (CT) images given initial expert user guidance, 2) validate the tool's accuracy and reliability in abdominal disease sites by comparison with manual segmentation by radiation oncology specialists, and 3) evaluate the tool's usability and time savings relative to manual segmentation. Initially the algorithm receives training information from the user via a set of intuitive brush strokes on one or a few 2D images, which yields statistics for setting parameters of a cost function that has both regional and boundary terms. The proposed method minimizes the cost function using a graph partition algorithm to achieve a segmentation that is propagated through an entire 3D stack of images. Our proposed further development will adapt the framework to true 3D segmentation to reduce the amount of user interaction and improve efficiency. A statistically large set of patient images manually segmented by experts will serve as ground truth for validation, which will analyze receiver-operator characteristics, volume overlap and surface distances between semi-automatic and manual segmentations. Feedback from radiation oncology specialists will provide evaluation of usability, while efficiency improvement will compare elapsed time between the semiautomatic tool and fully manual segmentation. The pilot study will focus on normal dose-limiting organs and structures in abdominal disease sites. RELEVANCE: The current practice in radiation treatment planning of cancer patients involves manual drawing of anatomical structures in computed tomography (CT) images, which is time consuming and introduces errors caused by user variability. We propose to develop and evaluate a software tool that will aid the expert user to rapidly and more accurately outline anatomical structures in CT images.
Hu YC, Grossberg MD, Mageras GS. Volumetric medical image segmentation with probabilistic conditional random fields framework. Proc 2009 Int Conf Image Processing, Computer Vision & Pattern Recognition, Las Vegas: CSREA Press. 2009; pp. 48-53.