Display options
Cite
Heredia G, Koay G, Turski P, et al. SU-D-218-01: Support Vector Machine Tissue Classification of Multiparametric MRI Tumor Data. Med Phys. 2012;39(6):3622doi: 10.1118/1.4734704.
Heredia, G., Koay, G., Turski, P., & Meyerand, M. (2012). SU-D-218-01: Support Vector Machine Tissue Classification of Multiparametric MRI Tumor Data. Medical physics, 39(6), 3622. https://doi.org/10.1118/1.4734704
Heredia, G, et al. "SU-D-218-01: Support Vector Machine Tissue Classification of Multiparametric MRI Tumor Data." Medical physics vol. 39,6 (2012): 3622. doi: https://doi.org/10.1118/1.4734704
Heredia G, Koay G, Turski P, Meyerand M. SU-D-218-01: Support Vector Machine Tissue Classification of Multiparametric MRI Tumor Data. Med Phys. 2012 Jun;39(6):3622. doi: 10.1118/1.4734704. PMID: 28517375.
Copy Download .nbib Format: NLM
Share it on

Med Phys. 2012 Jun;39(6):3622. doi: 10.1118/1.4734704.

SU-D-218-01: Support Vector Machine Tissue Classification of Multiparametric MRI Tumor Data.

Medical physics

G Heredia, G Koay, P Turski, M Meyerand

Affiliations

  1. University of Wisconsin, Madison, WI.

PMID: 28517375 DOI: 10.1118/1.4734704

Abstract

PURPOSE: To test the accuracy of Support Vector Machines in the classification of Glioblastoma Multiforme tumor voxels usingmultiparametric MRI data.

METHODS: Various MRI scans were collected from patients with recurrent GBM. Each scan session collected post-contrastT1(+C T1), T2, diffusion, perfusion, and multi-echo hypoxia images. Thediffusion-weighted images were converted to Apparent Diffusion Coefficient (ADC) maps. The perfusion images were corrected for leakage and represented as corrected rCBV maps. All of these scans were co-registered to each other, giving an input matrix to our support vector machine consisting of roughly 13,000 voxels, each with 5 feature values (+C T1, T2, ADC, rCBVcorrected, delta T2*). The SVM was then trained usingradiologist confirmed labels for 'cyst', 'tumor', and 'normal tissue'. These labels were obtained using longitudinal data as well as clinical scans, and tested on new data to determine the accuracy of classifying tumor and cyst voxels.

RESULTS: The tumor model resulted in a specificity of 0.9841 and a sensitivity of 0.7498. The cyst model resulted in a specificity of 0.9825 anda sensitivity of 0.9414. Both models showed improvement with increasing features.

CONCLUSIONS: These results show that the SVM is capable of classifying tumor and cyst voxels in a single case study. Aside from optimizing the current tumor model, future work will focus on the potential for the SVM to help in early detection of recurrence. In order to achieve this,we will need to test SVMs across patients. If we truly aim for early detection, then we need to prove that the algorithm can be trained on a pool of subjects with recurrent disease, and then test that model on a new patient before recurrence is obvious.

© 2012 American Association of Physicists in Medicine.

Keywords: Cancer; Diffusion; Magnetic resonance imaging; Medical imaging; Tissues

Publication Types

Grant support