Encrypted login | home

Program Information

Organ Specific Eigen-Texture for Atlas Based Segmentation

no image available
A Apte

A Apte*, S Fontenla , H Veeraraghavan , U Mahmood , J Oh , J Deasy , Memorial Sloan-Kettering Cancer Center, New York, NY


SU-F-FS4-7 (Sunday, July 30, 2017) 2:05 PM - 3:00 PM Room: Four Seasons 4

Purpose: Methodology for creating an atlas of eigen-textures that capture organ specific characteristics for atlas based image segmentation.

Methods: In this work, we propose deriving eigen-textures that capture the characteristics of the organs using eigen analysis of voxel-wise texture images. Eigen-texture images characterize voxel neighborhood using various types of textures such as Haralick and Laws features. The eigen-texture was computed from all the voxels in and around the organ of interest from a pre-selected number of patients. Once the eigen-texture is learned, the original images are transformed to create an atlas of images projected along the eigen-texture. In order to atlas-segment a new image, it has to be transformed using the same eigen transformation. The ROI for the test image can be selected by the user or automatically selected based on an initial segmentation that needs refinement.

Results: CT scans for 8 H&N cancer patients were used to build the atlas and to derive the eigen-texture. Parotid glands were the organ of interest. CERR software was used to compute the eigen-texture and commercial MIM software was used to store the atlas; and for atlas-based segmentation. The region of interest for generating the eigen-image for the test scan was based on an expanded bounding box around the clinical contour; whereas the entire scan was used for CT based atlas segmentation. Atlas based segmentation using eigen-textures matched closely with the clinically approved contour for the test image. The segmentation using the atlas of cropped, derived images outperformed the one using the atlas of full CT scans of the same patients.

Conclusion: Eigen-textures are potentially useful as atlases for image segmentation. The preliminary results indicate that an atlas derived from eigen-textures can be useful to refine an initial segmentation (for example, a bounding box).

Contact Email: