SU-C-18A-01: Online Atlas Selection Using 3D Gabor Features

Authors


Abstract

Purpose:

To develop and validate 3D Gabor features for on-the-fly selection of the best atlas to improve atlas-based segmentation.

Methods:

Twenty-one planning CT images of head and neck cancer patients were acquired and the submandibular glands and esophagus were contoured manually by a head-and-neck oncologist. One image was randomly chosen as the test and the remaining became the atlases. All atlases were rigidly registered to the test image and the union of the transformed atlas contours defined a local region for each organ. A bank of 3D Gabor filters was created with different scales and directions in the 3D space. These filters were applied to each image in the defined local region to extract 3D Gabor features, which are composed of the mean and standard deviation of each Gabor filtered 3D data. A distance metric was defined for Gabor features to determine the similarity between an atlas and the test image. For validation, the organ contours of all atlases were deformed to the test image and compared with the manual contour using Dice similarity coefficient, which determined the actual similarity.

Results:

A number of 144 Gabor filters (4 scales and 36 directions) was created to extract Gabor features for submandibular glands. The best atlas identified by the Gabor features was actually the second best atlas determined by the Dice value (71.7%); however, it is very close to the actual best atlas (73.3%). Similarly, a number of 48 Gabor filters (3 scales and 16 directions) was created for selecting the best esophagus atlas. The Gabor features identified the three best atlases exactly as they were determined by the Dice values.

Conclusion:

The 3D Gabor features showed great potential for online atlas selection to improve atlas-based segmentation.

This work was partially supported by a sponsored research grant from Varian Medical Systems.

Ancillary