<abstract xmlns="http://www.w3.org/1999/xhtml">

<p>The automatic segmentation of medical images stands at the basis of modern medical diagnosis, therapy planning and follow-up studies after interventions. The accuracy of the segmentation is a key element in assisting the work of the physician, but the efficiency of the process is also relevant. This paper introduces a feature selection strategy that attempts to define reduced feature sets for ensemble learning methods employed in brain tumor segmentation based on MRI data such a way that the segmentation outcome hardly suffers any damage. Initially, the full set of observed and generated features are deployed in ensemble training and prediction on testing data, which provide us information on all couples of features from the full feature set. The extracted pairwise data is fed to a Markov clustering (MCL) algorithm, which uses a graph structure to characterize the relation between features. MCL produces connected subgraphs that are totally separated from each other. The largest such subgraph defines the group of features which are selected for evaluation. The proposed technique is evaluated using the high-grade and low-grade tumor records of the training dataset of the BraTS 2019 challenge, in an ensemble learning framework relying on binary decision trees. The proposed method can reduce the set of features to 30%ofits initial size without losing anything in terms of segmentation accuracy, significantly contributing to the efficiency of the segmentation process. A detailed comparison of the full set of 104 features and the reduced set of 41 features is provided, with special attention to highly discriminative and redundant features within the MRI data.</p>
</abstract>

The automatic segmentation of medical images stands at the basis of modern medical diagnosis, therapy planning and follow-up studies after interventions. The accuracy of the segmentation is a key element in assisting the work of the physician, but the efficiency of the process is also relevant. This paper introduces a feature selection strategy that attempts to define reduced feature sets for ensemble learning methods employed in brain tumor segmentation based on MRI data such a way that the segmentation outcome hardly suffers any damage. Initially, the full set of observed and generated features are deployed in ensemble training and prediction on testing data, which provide us information on all couples of features from the full feature set. The extracted pairwise data is fed to a Markov clustering (MCL) algorithm, which uses a graph structure to characterize the relation between features. MCL produces connected subgraphs that are totally separated from each other. The largest such subgraph defines the group of features which are selected for evaluation. The proposed technique is evaluated using the high-grade and low-grade tumor records of the training dataset of the BraTS 2019 challenge, in an ensemble learning framework relying on binary decision trees. The proposed method can reduce the set of features to 30%ofits initial size without losing anything in terms of segmentation accuracy, significantly contributing to the efficiency of the segmentation process. A detailed comparison of the full set of 104 features and the reduced set of 41 features is provided, with special attention to highly discriminative and redundant features within the MRI data.

A feature selection strategy using Markov clustering, for the optimization of brain tumor segmentation from MRI data

Sapientia Hungarian University of Transylvania

Computational Intelligence Research Group, Sapientia Hungarian University of Transylvania

Acta Universitatis Sapientiae, Informatica

This work is licensed under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

The automatic segmentation of medical images stands at the basis of modern medical diagnosis, therapy planning and follow-up studies after interventions. The...

A feature selection strategy using Markov clustering, for the optimization of brain tumor segmentation from MRI data

Published Online: Feb 04, 2023

Page range: 316 - 337

Received: Dec 31, 2022

Accepted: Jan 18, 2023

DOI: https://doi.org/10.2478/ausi-2022-0018

Keywords
machine learning, feature selection, Markov clustering, image segmentation

© 2022 Ioan-Marius Pisak-Lukáts et al., published by Sciendo

This work is licensed under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

A feature selection strategy using Markov clustering, for the optimization of brain tumor segmentation from MRI data

Published Online: Feb 04, 2023

Page range: 316 - 337

Received: Dec 31, 2022

Accepted: Jan 18, 2023

DOI: https://doi.org/10.2478/ausi-2022-0018

Keywordsmachine learning, feature selection, Markov clustering, image segmentation

© 2022 Ioan-Marius Pisak-Lukáts et al., published by Sciendo

This work is licensed under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Keywords
machine learning, feature selection, Markov clustering, image segmentation