Program Information
Incorporating Mutual Information Into Atlas Selection for Multi-Atlas Based Rectum Auto-Contouring
L Qiu*, K Stephans , C Shah , A Godley , The Cleveland Clinic, Cleveland, OH
Presentations
MO-L-GePD-JT-4 (Monday, July 31, 2017) 1:15 PM - 1:45 PM Room: Joint Imaging-Therapy ePoster Theater
Purpose: To use mutual information to select the optimal atlases for multi-atlas based rectum auto-contouring.
Methods: 30 IGRT prostate patients, each with 20 to 28 daily kV CBCTs, had their bladder/prostate/rectum (B/P/R) contoured on each CBCTs. B/P/R contours were also automatically generated on the CBCTs by deformably registering the prior five days’ CBCTs (atlases) to the current CBCT and deforming the prior contours with the resulting deformation vector field. This set of deformed contours on the current CBCT is then combined using the STAPLE algorithm. Dice similarity coefficient (DSC) is used to evaluate the agreement between the auto and manual contours. The average DSCs of B/P/R auto-contours are 0.92/0.91/0.84. 28 CBCTs with poor rectum auto-contours were chosen for this study with average rectum DSC of only 0.778±0.087. We generated a rectum region of interest (ROI) contour for every target CBCT by subtracting the 1.5 mm contraction of the rectum auto contour from the 1.5 mm expansion of the same contour, leaving a 3 mm “rectal wall” contour. The mutual information (MI) from within this ROI was calculated to measure the accuracy of the deformation of the atlas to the current CBCT. Atlases with low MI were eliminated and STAPLE is rerun to create the new rectum auto contour.
Results: The MI of the ROI varied from 0.25 to 0.68. A poor DIR due to bad image quality or artifacts around the ROI generally leads to lower MI value. A strong correlation between MI and DSC of the rectum contour is observed. After removing atlases with low MI, the average DSC of the rectum auto-contours improves to 0.847±0.066 (p<0.0001 in paired t-test).
Conclusion: By using MI around the rectal wall to select atlases for auto-contouring, the accuracy of rectum contours is significantly improved.
Funding Support, Disclosures, and Conflict of Interest: A. Godley receives a research grant from Elekta AB, Stockholm, Sweden.
Contact Email: