On feature relevance in image-based prediction models: an empirical study

TY - GEN

T1 - On feature relevance in image-based prediction models

T2 - 4th International Workshop on Machine Learning in Medical Imaging

AU - Konukoglu, Ender

AU - Ganz, Melanie

AU - Van Leemput, Koen

AU - Sabuncu, Mert R.

N1 - Conference code: 4

PY - 2013

Y1 - 2013

N2 - Determining disease-related variations of the anatomy and function is an important step in better understanding diseases and developing early diagnostic systems. In particular, image-based multivariate prediction models and the "relevant features" they produce are attracting attention from the community. In this article, we present an empirical study on the relevant features produced by two recently developed discriminative learning algorithms: neighborhood approximation forests (NAF) and the relevance voxel machine (RVoxM). Specifically, we examine whether the sets of features these methods produce are exhaustive; that is whether the features that are not marked as relevant carry disease-related information. We perform experiments on three different problems: image-based regression on a synthetic dataset for which the set of relevant features is known, regression of subject age as well as binary classification of Alzheimer's Disease (AD) from brain Magnetic Resonance Imaging (MRI) data. Our experiments demonstrate that aging-related and AD-related variations are widespread and the initial sets of relevant features discovered by the methods are not exhaustive. Our findings show that by knocking-out features and re-training models, a much larger set of disease-related features can be identified.

AB - Determining disease-related variations of the anatomy and function is an important step in better understanding diseases and developing early diagnostic systems. In particular, image-based multivariate prediction models and the "relevant features" they produce are attracting attention from the community. In this article, we present an empirical study on the relevant features produced by two recently developed discriminative learning algorithms: neighborhood approximation forests (NAF) and the relevance voxel machine (RVoxM). Specifically, we examine whether the sets of features these methods produce are exhaustive; that is whether the features that are not marked as relevant carry disease-related information. We perform experiments on three different problems: image-based regression on a synthetic dataset for which the set of relevant features is known, regression of subject age as well as binary classification of Alzheimer's Disease (AD) from brain Magnetic Resonance Imaging (MRI) data. Our experiments demonstrate that aging-related and AD-related variations are widespread and the initial sets of relevant features discovered by the methods are not exhaustive. Our findings show that by knocking-out features and re-training models, a much larger set of disease-related features can be identified.

U2 - 10.1007/978-3-319-02267-3_22

DO - 10.1007/978-3-319-02267-3_22

M3 - Article in proceedings

SN - 978-3-319-02266-6

T3 - Lecture notes in computer science

SP - 171

EP - 178

BT - Machine Learning in Medical Imaging

A2 - Wu, Guorong

A2 - Zhang, Daoqiang

A2 - Shen, Dinggang

A2 - Yan, Pingkun

A2 - Suzuki, Kenji

A2 - Wang, Fei

PB - Springer Publishing Company

Y2 - 22 September 2013 through 22 September 2013

ER -