Feature-space transformation improves supervised segmentation across scanners

Annegreet van Opbroek; Hakim C. Achterberg; Marleen de Bruijne

doi:10.1007/978-3-319-27929-9_9

Feature-space transformation improves supervised segmentation across scanners

Annegreet van Opbroek^*, Hakim C. Achterberg, Marleen de Bruijne

^*Corresponding author for this work

Department of Computer Science

3 Citations (Scopus)

Abstract

Image-segmentation techniques based on supervised classification generally perform well on the condition that training and test samples have the same feature distribution. However, if training and test images are acquired with different scanners or scanning parameters, their feature distributions can be very different, which can hurt the performance of such techniques. We propose a feature-space-transformation method to overcome these differences in feature distributions. Our method learns a mapping of the feature values of training voxels to values observed in images from the test scanner. This transformation is learned from unlabeled images of subjects scanned on both the training scanner and the test scanner. We evaluated our method on hippocampus segmentation on 27 images of the Harmonized Hippocampal Protocol (HarP), a heterogeneous dataset consisting of 1.5T and 3T MR images. The results showed that our feature space transformation improved the Dice overlap of segmentations obtained with an SVM classifier from 0.36 to 0.85 when only 10 atlases were used and from 0.79 to 0.85 when around 100 atlases were used.

Original language	English
Title of host publication	Machine learning meets medical imaging : First International Workshop, MLMMI 2015, Held in Conjunction with ICML 2015, Lille, France, July 11, 2015, Revised Selected Papers
Number of pages	9
Publisher	Springer
Publication date	2015
Pages	85-93
DOIs	https://doi.org/10.1007/978-3-319-27929-9_9
Publication status	Published - 2015
Event	1st International Workshop on Machine Learning Meets Medical Imaging, MLMMI 2015 - Lille, France Duration: 11 Jul 2015 → 11 Jul 2015

Conference

Conference	1st International Workshop on Machine Learning Meets Medical Imaging, MLMMI 2015
Country/Territory	France
City	Lille
Period	11/07/2015 → 11/07/2015

Series	Lecture notes in computer science
Volume	9487
ISSN	0302-9743

Keywords

Brain
Hippocampus
Machine learning
MRI
Transfer learning

Access to Document

10.1007/978-3-319-27929-9_9

Cite this

van Opbroek, A., Achterberg, H. C., & de Bruijne, M. (2015). Feature-space transformation improves supervised segmentation across scanners. In Machine learning meets medical imaging: First International Workshop, MLMMI 2015, Held in Conjunction with ICML 2015, Lille, France, July 11, 2015, Revised Selected Papers (pp. 85-93). Springer. Lecture notes in computer science Vol. 9487 https://doi.org/10.1007/978-3-319-27929-9_9

Feature-space transformation improves supervised segmentation across scanners. / van Opbroek, Annegreet; Achterberg, Hakim C.; de Bruijne, Marleen.

Machine learning meets medical imaging: First International Workshop, MLMMI 2015, Held in Conjunction with ICML 2015, Lille, France, July 11, 2015, Revised Selected Papers. Springer, 2015. p. 85-93 (Lecture notes in computer science, Vol. 9487).

Research output: Chapter in Book/Report/Conference proceeding › Article in proceedings › Research › peer-review

van Opbroek, A, Achterberg, HC & de Bruijne, M 2015, Feature-space transformation improves supervised segmentation across scanners. in Machine learning meets medical imaging: First International Workshop, MLMMI 2015, Held in Conjunction with ICML 2015, Lille, France, July 11, 2015, Revised Selected Papers. Springer, Lecture notes in computer science, vol. 9487, pp. 85-93, 1st International Workshop on Machine Learning Meets Medical Imaging, MLMMI 2015, Lille, France, 11/07/2015. https://doi.org/10.1007/978-3-319-27929-9_9

van Opbroek A, Achterberg HC, de Bruijne M. Feature-space transformation improves supervised segmentation across scanners. In Machine learning meets medical imaging: First International Workshop, MLMMI 2015, Held in Conjunction with ICML 2015, Lille, France, July 11, 2015, Revised Selected Papers. Springer. 2015. p. 85-93. (Lecture notes in computer science, Vol. 9487). doi: 10.1007/978-3-319-27929-9_9

van Opbroek, Annegreet ; Achterberg, Hakim C. ; de Bruijne, Marleen. / Feature-space transformation improves supervised segmentation across scanners. Machine learning meets medical imaging: First International Workshop, MLMMI 2015, Held in Conjunction with ICML 2015, Lille, France, July 11, 2015, Revised Selected Papers. Springer, 2015. pp. 85-93 (Lecture notes in computer science, Vol. 9487).

@inproceedings{361ffa3cabb94484be23f2b79267bd45,

title = "Feature-space transformation improves supervised segmentation across scanners",

abstract = "Image-segmentation techniques based on supervised classification generally perform well on the condition that training and test samples have the same feature distribution. However, if training and test images are acquired with different scanners or scanning parameters, their feature distributions can be very different, which can hurt the performance of such techniques. We propose a feature-space-transformation method to overcome these differences in feature distributions. Our method learns a mapping of the feature values of training voxels to values observed in images from the test scanner. This transformation is learned from unlabeled images of subjects scanned on both the training scanner and the test scanner. We evaluated our method on hippocampus segmentation on 27 images of the Harmonized Hippocampal Protocol (HarP), a heterogeneous dataset consisting of 1.5T and 3T MR images. The results showed that our feature space transformation improved the Dice overlap of segmentations obtained with an SVM classifier from 0.36 to 0.85 when only 10 atlases were used and from 0.79 to 0.85 when around 100 atlases were used.",

keywords = "Brain, Hippocampus, Machine learning, MRI, Transfer learning",

author = "{van Opbroek}, Annegreet and Achterberg, {Hakim C.} and {de Bruijne}, Marleen",

year = "2015",

doi = "10.1007/978-3-319-27929-9_9",

language = "English",

series = "Lecture notes in computer science",

publisher = "Springer",

pages = "85--93",

booktitle = "Machine learning meets medical imaging",

note = "1st International Workshop on Machine Learning Meets Medical Imaging, MLMMI 2015 ; Conference date: 11-07-2015 Through 11-07-2015",

}

TY - GEN

T1 - Feature-space transformation improves supervised segmentation across scanners

AU - van Opbroek, Annegreet

AU - Achterberg, Hakim C.

AU - de Bruijne, Marleen

PY - 2015

Y1 - 2015

N2 - Image-segmentation techniques based on supervised classification generally perform well on the condition that training and test samples have the same feature distribution. However, if training and test images are acquired with different scanners or scanning parameters, their feature distributions can be very different, which can hurt the performance of such techniques. We propose a feature-space-transformation method to overcome these differences in feature distributions. Our method learns a mapping of the feature values of training voxels to values observed in images from the test scanner. This transformation is learned from unlabeled images of subjects scanned on both the training scanner and the test scanner. We evaluated our method on hippocampus segmentation on 27 images of the Harmonized Hippocampal Protocol (HarP), a heterogeneous dataset consisting of 1.5T and 3T MR images. The results showed that our feature space transformation improved the Dice overlap of segmentations obtained with an SVM classifier from 0.36 to 0.85 when only 10 atlases were used and from 0.79 to 0.85 when around 100 atlases were used.

AB - Image-segmentation techniques based on supervised classification generally perform well on the condition that training and test samples have the same feature distribution. However, if training and test images are acquired with different scanners or scanning parameters, their feature distributions can be very different, which can hurt the performance of such techniques. We propose a feature-space-transformation method to overcome these differences in feature distributions. Our method learns a mapping of the feature values of training voxels to values observed in images from the test scanner. This transformation is learned from unlabeled images of subjects scanned on both the training scanner and the test scanner. We evaluated our method on hippocampus segmentation on 27 images of the Harmonized Hippocampal Protocol (HarP), a heterogeneous dataset consisting of 1.5T and 3T MR images. The results showed that our feature space transformation improved the Dice overlap of segmentations obtained with an SVM classifier from 0.36 to 0.85 when only 10 atlases were used and from 0.79 to 0.85 when around 100 atlases were used.

KW - Brain

KW - Hippocampus

KW - Machine learning

KW - MRI

KW - Transfer learning

U2 - 10.1007/978-3-319-27929-9_9

DO - 10.1007/978-3-319-27929-9_9

M3 - Article in proceedings

AN - SCOPUS:84955254180

T3 - Lecture notes in computer science

SP - 85

EP - 93

BT - Machine learning meets medical imaging

PB - Springer

T2 - 1st International Workshop on Machine Learning Meets Medical Imaging, MLMMI 2015

Y2 - 11 July 2015 through 11 July 2015

ER -

Feature-space transformation improves supervised segmentation across scanners

Abstract

Conference

Keywords

Access to Document

Fingerprint

Cite this