Automatic Pelvic Organ Delineation in Prostate Cancer Treatment Abstract: Fast, reliable and accurate delineation of pelvic organs in the planning and treatment images is a long- standing, important and technically challenging problem. Its solution is highly required for state-of-the-art image-guided radiation therapy planning and treatment, as better treatment decisions rely on timely interpretation of anatomical information in the images. However, automatic segmentation in male pelvic regions is always difficult due to 1) low contrast between prostate and surrounding organs, and 2) possibly disparate shapes/appearances of bladder and rectum caused by tissue deformations. The goal of this project is to create a set of novel machine learning tools to achieve accurate, reliable and efficient delineation of important pelvic organs (e.g., prostate, bladder, and rectum) in different modalities (e.g., planning CT, treatment CT/CBCT, and MRI) for radiotherapy of prostate cancer. Planning CT. For automatic segmentation, landmark detection is often the first step in rapidly locating the target organs. Thus, in Aim 1, we will create a novel joint landmark detection approach, based on both random forests and auto-context model, to iteratively detect all landmarks and further coordinate their detection results for achieving more accurate and consistent landmark detection results. After roughly locating organs with the aid of those detected landmarks, the second step is to accurately segment boundaries of target organs in the planning CT. Accordingly, in Aim 2, we will create a set of learning methods to a) first simultaneously predict all pelvic organ boundaries in the planning CT with the regression forests trained by labeled training data, and b) then segment all pelvic organs jointly by deforming their respective shape models. In particular, to address the limitations of conventional deformable models in assuming simple Gaussian distributions for organ shapes, a novel hierarchical sparse shape composition approach will be developed to constrain shape models during deformable segmentation. Treatment CT/CBCT. During the course of serial radiation treatments, to quantitatively record and monitor the accumulated dose delivered to the patient, organs in the treatment image also need to be segmented. Although methods proposed in Aims 1-2 can be simply applied, as done by many conventional methods, this will lead to a) inconsistent landmark detection and b) inconsistent segmentations across different treatment days because of possible large shape/appearance changes. Accordingly, in Aim 3, we will create a novel self- learning mechanism to gradually learn and incorporate patient-specific information into both joint landmark detection and deformable segmentation steps from the increasingly acquired treatment images of patient. Thus, population data will gradually be replaced by the patient's own data to train personalized models. MRI. To guide pelvic organ seg...