The use of LS-SVM in the classification of brain tumors based on Magnetic Resonance Spectroscopy signals

Lukas Lukas, Andy Devos, Johan A. K. Suykens, Leentje Vanhamme, Sabine Van Huffel, Anne Rosemary Tate, Carles Majós, Carles ArúsESANN 2002

Abstract

Accurate classification of brain tumors is vital for determining treatment pathways, yet traditional biopsies are invasive and carry risks. This paper explores the application of Least Squares Support Vector Machines (LS-SVM) to classify brain tumors using non-invasive Magnetic Resonance Spectroscopy (MRS) data. By analyzing the chemical signatures of tissue, the researchers developed a model capable of distinguishing between glioblastomas and meningiomas with high accuracy. Despite a small dataset of 47 spectra, the LS-SVM with a Radial Basis Function (RBF) kernel achieved a classification accuracy of 91.5% and an Area Under the Curve (AUC) of 0.9438, demonstrating the potential for automated, metabolite-based diagnostic tools.

Introduction

Identifying the specific type and grade of a brain tumor is a critical clinical challenge. While Magnetic Resonance Imaging (MRI) provides essential anatomical context, Magnetic Resonance Spectroscopy (MRS) offers a deeper look into the tissue's underlying biochemistry. MRS signals represent the concentration of various metabolites, creating a "chemical fingerprint" of the tumor. However, these signals are high-dimensional and datasets are often small, leading to the "curse of dimensionality" where standard statistical methods fail. This study investigates whether LS-SVMs can effectively navigate these complex datasets to provide robust, automated tumor discrimination.

Methodology

The researchers utilized the Least Squares Support Vector Machine (LS-SVM), a variation of the standard SVM that simplifies the optimization process into a set of linear equations. This makes the algorithm computationally efficient and particularly suited for high-dimensional data.

Data Preprocessing and Feature Selection

The study used 47 MR spectra from the INTERPRET project (23 glioblastomas and 24 meningiomas). The raw time-domain signals underwent rigorous preprocessing, including phase correction, water peak filtering via the HLSVD algorithm, and transformation into the frequency domain using Fast Fourier Transform (FFT). The final feature set focused on a specific frequency range (-34 Hz to -240 Hz) containing 107 variables representing key metabolites such as choline (Cho), creatine (Cr), and lactate (Lac).

Kernel Selection

To find the optimal mapping for tumor separation, three kernel types were compared:

  1. Linear Kernel
  2. Polynomial Kernel
  3. Radial Basis Function (RBF) Kernel

Chemical Fingerprinting and Insights

A key finding of the study is the distinct metabolic profiles associated with different tumor types. Glioblastomas (Class 1) are often characterized by prominent lipid and lactate peaks, whereas meningiomas (Class 2) frequently exhibit unique alanine (Ala) signals.

Comparison of mean MRS spectra for glioblastomas (left) and meningiomas (right), highlighting key metabolite differences like lipids/Lac and Ala.

The model's success suggests that these chemical signatures are sufficiently distinct for automated classification even without supplementary clinical data like patient age or gender.

Results and Performance

The performance of the LS-SVM models was evaluated using leave-one-out (LOO) cross-validation. The RBF kernel significantly outperformed the linear and polynomial alternatives, suggesting that the relationship between metabolite concentrations and tumor pathology is inherently non-linear.

Kernel Type Accuracy AUC
RBF 91.5% 0.9438
Polynomial 78.7% 0.8170
Linear 78.7% 0.7862
ROC curves for the three kernel types, showing the superior discriminative power of the RBF kernel (top-most curve).

The RBF kernel's AUC of 0.9438 indicates an excellent ability to discriminate between the two tumor classes, providing a substantial improvement over simpler linear models.

Conclusion

This research demonstrates that LS-SVMs, particularly when paired with an RBF kernel, are highly effective at classifying brain tumors from MRS data. While the sample size was limited to 47 patients, the high accuracy and AUC scores are promising. Future work will focus on expanding the dataset to improve model generalization and integrating MRS signals with other clinical imaging modalities to create a comprehensive, non-invasive diagnostic platform.

Summary updated

BibTeX
@inproceedings{lukas2002use,
  author = {Lukas Lukas and Andy Devos and Johan A. K. Suykens and Leentje Vanhamme and Sabine Van Huffel and Anne Rosemary Tate and Carles Majós and Carles Arús},
  title = {{The use of LS-SVM in the classification of brain tumors based on Magnetic Resonance Spectroscopy signals}},
  booktitle = {ESANN},
  year = {2002},
}
Figure