University of North Carolina Comparison of Human and M-rep Kidneys Segmented from CT Images James Chen, Gregg Tracton, Manjari Rao, Sarang Joshi, Steve Pizer, Joshua Stough, and Ed Chaney University of North Carolina Chapel Hill
Methods [ Target Images: 12 planning CT images from archives (24 kidneys) [ Human (A and B) segmentation: Careful slice-by-slice pixel painting (voxel-based) [ Computer segmentation: Deformable m-reps (yields smooth surface) [ Comparison metrics reported here: Volume overlap and mean surface separation
University of North Carolina The median volume overlap is measured by the overlap volume divided by the union of the two volumes being compared.
Image Challenges Poor contrast (no contrast agent) Crowded soft tissue environment
Image Challenges Motion artifacts
Image Challenges Image variability near renal pelvis Three human segmentations of renal pelvis
Three Stages of M-rep Segmentation Similarity + elongation Atom deformation Boundary displacement Implied surface after each scale of segmentation
University of North Carolina Result after Final Stage
Limitations of Comparison [ Segmented structures represented as collections of whole 2 mm 3 voxels [ Conversion of m-rep surfaces to voxels introduces a bias that favors humans [ Sensitivity of distance ~ 2 mm [ Best possible volume overlap is ~ 95%
Results: Volume Overlap Human A to Human B: 90%-96% M-rep to Human A or B: 85%-95%
University of North Carolina Case Human A vs B volume overlap Kidney Number
University of North Carolina M-rep vs Human A volume overlap Case
University of North Carolina
Results: Average Surface Separation Human A to Human B: Within one voxel M-rep to Human A or B: Within one voxel * * with one exception
University of North Carolina Image Image Human Human M-rep M-rep Outlier Case
Who is right (Human or M-rep)? Movie Loop
Conclusions In this study m-reps compared with humans A and B as well as A compared with B.
University of North Carolina Sunset at Portsmouth Island