SU-F-I-42: Evaluation of Multiple Atlas Segmentation Methods for the Abdomen.

PURPOSE: Atlas segmentation has the potential to provide time savings for contour generation in various treatment sites. Previously we evaluated the optimal number of atlas matches for head and neck, lung, liver, and prostate cancer. Our goal in the current work is to find the optimal number of atlas matches and method for combining atlas contours for an abdominal CT atlas.

METHODS: An abdominal CT atlas containing 37 subjects was utilized for atlas segmentation and was evaluated using a leave-one-out methodology. Each atlas subject contained manually defined contours of the aorta, left and right kidneys, large bowel, liver, spinal cord, stomach, and vertebral body. The one, three, four, five, and seven best matches from the atlas were automatically found for each subject and were deformed to the test subject. The contours were combined for each multi-atlas method using Majority Vote (MV) and STAPLE. The average Dice Similarity Index was calculated for each method compared to manual contours.

RESULTS: Majority vote with five (MV-5) and seven matches (MV-7) had the highest dice for 6 out of the 8 structures including the aorta (0.75), left kidney (0.80), right kidney (0.83 and 0.84), liver (0.91), spinal cord (0.78), and vertebral body (0.71 and 0.72). For the large bowel, STAPLE-5 and MV-4 had the highest dice score (0.38 and 0.37), while MV4 provided the highest dice (0.65) for the stomach. The best method was statistically significant over the next best method for the aorta, left and right kidneys, liver, large bowel, and stomach with p-values ranging from 0.0002 - 0.05.

CONCLUSION: Atlas-based segmentation using MV-5 and MV-7 provided the most accurate contours for the majority of structures, with only STAPLE-5 and MV-4 being more accurate for the large bowel and MV-4 being more accurate for the stomach. Aaron Nelson is a part owner and employee of MIM Software Inc. Sara Pirozzi is an employee of MIM Software Inc.