SU‐FF‐I‐86: Esophagus Segmentation in Thoracic CT Images for Radiotherapy Planning

• Published in: Medical Physics Vol. 36; no. 6; p. 2454
• Main Authors: Kurugol, S, Sharp, G, Dy, J, Brooks, D
• Format: Conference Proceeding Journal Article
• Language: English
• Published: American Association of Physicists in Medicine 01.06.2009
• Subjects: Anatomy; Computed tomography; Medical image segmentation; Medical imaging; Polynomials; Radiation therapy; Spatial analysis; Spatial dimensions; Spatial scaling; Trajectory models
• ISSN: 0094-2405; 2473-4209
• DOI: 10.1118/1.3181206

Purpose: Contouring organs at risk in thoracic CT images for radiotherapy planning is labor intensive. We propose a technique to semiautomatically segment the esophagus in thoracic CT images. Method and Materials: We used training datasets to learn the distribution of relative spatial location of esophagus centers with respect to neighboring structures and to construct models of their anterior‐posterior (x) and left‐right (y) coordinate trajectories as a function of cranial‐caudal (z) position after local normalization by transverse scaling and translation. In the z direction, we selected 8 anatomical reference points and matched these points for each training set. We observed that the centers can be modeled by a single cubic polynomial in the sagittal plane and by a shape‐preserving spline in the coronal plane. To segment the esophagus in a 3D dataset, we estimated the center based on histograms for each reference slice and fit the polynomial center models for other slices. We then used level‐sets to locate the esophagus wall. We initialized the level‐set function and shape prior using the estimated centers and updated to minimize an energy functional combining several existing region, shape and smoothness terms to impose a smooth slice‐to‐slice change of ellipse parameters. Results: Testing on 8 subjects against expert segmentations, the center‐estimation algorithm achieved 2mm average error in the x‐direction but was less accurate in the y‐direction with a 4.3mm average error. The level‐set method improved the average error in y‐direction to 2.8mm. Conclusion: Segmentation of the esophagus using prior information from training including spatial dependence between neighboring structures and models of the esophagus centers can be performed with a level set approach. Using spatial information and slice‐to‐slice smoothing improves the performance for regions with no contrast. Acknowledgement: The work of first and fourth author was supported by NIH/NCRR Center for Integrative Biomedical Computing (CIBC), P41‐RR12553‐09.