Please use this identifier to cite or link to this item:
|Title:||Atlas Selection for Automated Segmentation of Pelvic CT for Prostate Radiotherapy|
|Abstract:||Radiotherapy has become a standard modality for treating prostate cancer. Typically, intensity modulated radiotherapy (IMRT) is employed. Accurate delineation is important to ensure that the clinical target volume (prostate) is sufficiently irradiated and that the organs at risk (OARs) are appropriately spared. A recent technological development in radiotherapy treatment planning is the employment of atlas-based segmentation to automate target volume and OAR delineation. Atlas based-segmentation utilizes the spatial relationship between a pre-contoured atlas subject and a new patient image to derive the segmentation result. The typical approach has three steps: many atlas subject images are globally registered to the target image, an atlas subject image that is the most similar to the target is selected, and the chosen atlas image and contours are aligned with the target image space using deformable registration. The purpose of this work was to design an atlas selection strategy and evaluate its impact on the final atlas-based segmentation outcome. Segmentation accuracy was mainly quantified using the Dice Similarity Coefficient (DSC), which was used to score the overlap between automatic and manual contours on a 0 to 1 scale. An alternative atlas selection approach was proposed that identified the most similar atlas subject based on several anatomical measurements that were chosen to indicate the overall prostate and body shape. A brute force procedure was first performed for a training dataset of 20 patients using image registration to pair subjects with similar contours based on DSC. For the identified best matches, anatomical measurements were compared. An atlas selection procedure was designed; relying on the computation of a similarity score defined as a weighted sum of differences between the target and atlas subject anatomical measurements. Finally, an optimization procedure was performed to obtain the weights that gave the highest DSC between automatic and manual contours for the training set. The mean DSCs obtained using brute force were 0.78±0.07 and 0.90±0.02 for the prostate and either femoral head. The proposed atlas selection method achieved 0.72±0.11 and 0.87±0.03 for the prostate and either femoral head. Clearly, the algorithm was able to identify the best matching atlas subject for any target subject in the training set of data. The key point of this work was to also validate the atlas selection strategy. Thus, the optimized atlas selection procedure was tested on images of 10 additional subjects. Again, the algorithm’s ability to predict the most similar atlas subject was excellent. A brute force search for the set of 10 images achieved mean DSCs of 0.76±0.03 and 0.88±0.03 for the prostate and either femoral head. The proposed method yielded DSCs of 0.64±0.09 and 0.86±0.04 for the prostate and either femoral head. The difference in mean DSCs between the proposed method and the brute force was statically significant (p < 0.05). Overall, the brute force results demonstrated that atlas-based segmentation can reproduce a similar level of accuracy as manual recontouring (Granberg et al., 2011). More importantly, the same level of accuracy was achieved with the proposed atlas selection method as with more computationally intensive techniques. These results indicate that atlas-based segmentation is a promising technique for prostate radiation therapy.|
|Appears in Collections:||Open Access Dissertations and Theses|
Files in This Item:
|Atlas_Selection_for_Automated_Segmentation_of_Pelvic_CT_for_Prostate_Radiotherapy_thesis_Final.pdf||1.93 MB||Adobe PDF||View/Open|
Items in MacSphere are protected by copyright, with all rights reserved, unless otherwise indicated.