SU‐GG‐I‐89: Automatic Segmentation of Risk Structures

Abstract

Purpose: The clinical implementation of conformal radiotherapy IMRT has substantially increased the need for automated contouring tools. Implementation of image‐guided adaptive RT will further increase the need. The objective of this work is to develop and assess a fully automated method for delineating risk structures on the basis of statistical models and anatomical atlases. Method and Materials: Anatomical variability is modeled from a set of training images from multiple patients, where structures of interest were manually contoured by clinical experts, and salient anatomical landmarks were defined, e.g. the mastoid process. We use landmark correspondences and thin‐plate spline (TPS) approximation to map all training data into a common geometrical reference frame and generate a mean volumetric image and mean 3D surfaces; spatial variation and dependency of the landmarks is encoded in a statistical model and their local appearance is derived from the mean image. This model is then used to automatically find the positions of the landmarks in a new patient image. TPS is used to map the contours from the mean volume onto the new volume. Finally, the mapped contours are refined using a model‐based segmentation technique. Results: We evaluated our method for three anatomical regions: head‐neck, pelvis and thorax, on multiple clinical datasets accompanied with manually created contours. Quantitative results where collected for spatial accuracy, both for detection of the landmarks as well as for the final contours. For initialization the distance to ground‐truth was 8.9 ± 4.8 mm, contour distance was 3.5 ± 4.5 mm across all anatomies. Computation time was below 50 seconds in all cases. Conclusion: A fully automated segmentation method was developed and quantitatively evaluated for risk structures from CT images in the head‐neck, pelvis and thorax. Future work will increase consistency and efficiency for structures with poor image contrast.