MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Bagged Atlas Based Segmentation (BABS): Bagging Atlas Segmentations From Multiple Image Representations

A Apte1*, A Iyer1 , L V.van Dijk3 , M Thor1 , H Veeraraghavan1 , S Fontenla1 , J Oh1 , J Mechalakos1 , M Hunt1 , J Deasy1 , (1)Memorial Sloan Kettering Cancer Center, New York, NY, (2) University of Groningen, Groningen, Netherlands

Presentations

(Sunday, 7/29/2018) 4:00 PM - 4:55 PM

Room: Karl Dean Ballroom C

Purpose: To develop a novel paradigm, BABS, for image segmentation by bagging segmentations from various image representations that emphasize organ specific characteristics and combine with atlas fusion techniques.

Methods: The proposed segmentation paradigm is based on Principal Component (PC) Analysis of texture representations of CT images at voxels within regions of interest (ROIs) contoured by an expert. Based on the generated PCs, which emphasize characteristics such as bones and edges in and around the ROI, a new set of CT images were decomposed into the first two “eigen-texture� images. This resulted in a total of three atlases: (1) Original CT image, (2) The first eigen texture image (PC1) and (3) The second eigen texture image (PC2). To fuse segmentations from the atlas onto new patients, the following three techniques were used: (1) Majority vote, (2) STAPLE and (3) Generalized Registration Error (GRE) weighted score. The three image representations along with the three atlas fusion techniques resulted in nine different segmentations, which were further fused into the final segmentation using STAPLE.

Results: Principal components were derived from individual voxels within an expanded bounding box around the left parotid gland on CT images of 9 patients. A different set of CT scans for 24 head and neck cancer patients was used to evaluate the algorithm performance in a Leave-One-Out fashion. Left and right parotids segmented by the same expert were used as the ground truth. The improvement in Dice and 95th percentile Haussdorff metrics resulting from the BABS algorithm were statistically significant with Wilcoxon signed-rank test (p-values of 10-7 and 0.005 respectively) compared to the atlas segmentation using only the original CT images.

Conclusion: BABS resulted in an improved performance for segmenting parotid glands compared to the conventional multi-atlas segmentation using CT images.

Funding Support, Disclosures, and Conflict of Interest: This research was partially funded by NIH grant 1R01CA198121 and NIH/NCI Cancer Center Support grant P30 CA008748.

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email