Shape-based averaging for combination of multiple segmentations

Authors:
T. Rohlfing;C. R. Maurer, Jr.
Affiliations:
Neuroscience Program, SRI International, Menlo Park, CA;Department of Neurosurgery, Stanford University, Stanford CA
Venue:
MICCAI'05 Proceedings of the 8th international conference on Medical image computing and computer-assisted intervention - Volume Part II
Year:
2005

Citing 4
Cited 2

Free-form deformation of solid geometric models

SIGGRAPH '86 Proceedings of the 13th annual conference on Computer graphics and interactive techniques
On Combining Classifiers

IEEE Transactions on Pattern Analysis and Machine Intelligence
A Linear Time Algorithm for Computing Exact Euclidean Distance Transforms of Binary Images in Arbitrary Dimensions

IEEE Transactions on Pattern Analysis and Machine Intelligence
Multi-classifier framework for atlas-based image segmentation

Pattern Recognition Letters

Combination of Multiple Segmentations by a Random Walker Approach

Proceedings of the 30th DAGM symposium on Pattern Recognition
Revisiting the evaluation of segmentation results: introducing confidence maps

MICCAI'07 Proceedings of the 10th international conference on Medical image computing and computer-assisted intervention

Quantified Score

Hi-index	0.00

Visualization

Abstract

Combination of multiple segmentations has recently been introduced as an effective method to obtain segmentations that are more accurate than any of the individual input segmentations. This paper introduces a new way to combine multiple segmentations using a novel shape-based averaging method. Individual segmentations are combined based on the signed Euclidean distance maps of the labels in each input segmentation. Compared to label voting, the new combination method produces smoother, more regular output segmentations and avoids fragmentation of contiguous structures. Using publicly available segmented human brain MR images (IBSR database), we perform a quantitative comparison between shape-based averaging and label voting by combining random segmentations with controlled error magnitudes and known ground truth. Shape-based averaging generated combined segmentations that were closer to the ground truth than combinations from label voting for all numbers of input segmentations (up to ten). The relative advantage of shape-based averaging over voting was larger for fewer input segmentations, and larger for greater deviations of the input segmentations from the ground truth. We conclude that shape-based averaging improves the accuracy of combined segmentations, in particular when only a few input segmentations are available and when the quality of the input segmentations is low.