Program Information
Coupling Database Search with Atlas-Based Segmentation: Application to Esophageal Delineation
E Schreibmann* E Elder, Department of Radiation Oncology, Winship Cancer Institute of Emory University
Presentations
SU-E-T-228 (Sunday, July 12, 2015) 3:00 PM - 6:00 PM Room: Exhibit Hall
Purpose: Atlas segmentation is a popular method of warping a pre-segmented template to a new patient anatomy, however for anatomy in the thorax and abdomen large inter-patient variations exists, and a single template does not provide the required clinical accuracy. With the proposed method the CT dataset to be segmented is matched to a similar case in a database of pre-segmented templates.
Method: The key aspect of the approach is to search in a database of images and associated segmentations for a case with similar anatomy. A set of 29 previous cases have been pre-registered to a common reference and cropped to the esophageal region to construct a compact representation of the database through quantification with a principal component analysis algorithm that is designed to find similar cases by direct comparison of the eigenvectors for ranking database cases according to the similarity with the query case. Once a similar case at image level is identified in the database, an affine registration is performed to transfer the pre-segmentations from the database to the images to be segmented.
Results: The algorithm was applied for segmenting the esophageal region on 12 test cases and the conformity between the atlas and manual segmentations were measured using the Dice coefficient and the Euclidian distance. Dice coefficients ranged between (0.60, 0.87) with mean of 0.78 and standard deviation of 0.09. Mean distance ranged between (0.18,1.66) mm mean of 0.82 mm, SD of 0.51. Once registered to the database, searching in the database takes seconds as it does not compare the images directly but just the components of the PCA vectors, adding little overhead to the overall delineation algorithm.
Conclusions: Selection of a similar atlas from a database can be achieved with little computational overhead by using a PCA-based similarity metric as demonstrated for esophageal segmentation.
Contact Email: