Program Information
Surrogate Function Evaluation and Comparison for Atlas Selection in Multi-Atlas-Based Segmentation
Q Xie, D Ruan*, UCLA School of Medicine, Los Angeles, CA
Presentations
SU-C-18C-1 Sunday 1:00PM - 1:55PM Room: 18CPurpose: To develop a systematic surrogate evaluation and comparison scheme to select a subset of training samples that best resemble the target in multi-atlas-based automatic segmentation, for the purpose of improved accuracy.
Methods: In multi-atlas-based segmentation, it is desirable to fuse labels from a subset of training samples among the atlases that contribute positively to the final segmentation, for improved accuracy and efficiency. A surrogate metric is needed to prognosticate the contribution of a training image to the target segmentation. We formalize atlas subset selection as a ranking problem and develop a quality index to characterize the efficacy of a surrogate approach. A candidate surrogate is assessed in a supervised learning setting, where images with known segmentation are split into training atlases and testing targets, to enable assess to ground-truth contribution level from a training atlas to the target, as measured by the Dice similarity coefficient (DSC). The quality index measures the discrepancy between (1) DSC in descending order as from an Oracle and (2) the reordered DSCs based on the surrogate metric. The quality index was applied in the context of automatic prostate and brain segmentations, to assess 6 surrogate candidates, namely, mean square error (MSE), normalized cross correlation (NCC), and mutual information (MI), either measured within a fixed region of interest (ROI), fROI, or adaptive ROI (aROI).
Results: Surrogate metrics measured within aROI significantly outperform those within fROI (paired t-test, p-values consistently below 0.067), yet no statistically significant difference among MSE/NCC/MI (paired t-test, 0.014≤p≤0.941).
Conclusion: We have developed a quality index to quantitatively evaluate and compare surrogates' efficacy in atlas selection for automatic segmentation. When applied to comparing MSE/NCC/MI within fROI/aROI, it shows that surrogate metrics over aROI instead of fROI significantly improve segmentation performance, while there is no statistical evidence to support the superiority among MSE/NCC/MI.
Contact Email: