Manifold Learning in MR spectroscopy using nonlinear dimensionality reduction and unsupervised clustering.

Area of Science:

• Neuroimaging
• Biophysics
• Machine Learning

Background:

• Accurate classification of brain tumors using magnetic resonance spectroscopy (MRS) data is crucial for diagnosis and treatment planning.
• Unsupervised learning methods are increasingly applied to analyze complex MRS datasets, but their performance can be limited by data dimensionality.

Purpose of the Study:

• To evaluate the efficacy of nonlinear dimensionality reduction techniques, specifically Laplacian eigenmaps (LE), in improving unsupervised classification of (1) H MRS brain tumor data compared to linear methods.
• To assess the performance of LE in glioma grading and tissue type segmentation using both single-voxel (1) H MRS and (1) H magnetic resonance spectroscopy imaging (MRSI) data.

Main Methods:

• Acquired in vivo single-voxel (1) H MRS and (1) H MRSI data from 55 and 29 glioma patients, respectively.
• Applied Laplacian eigenmaps (LE) and independent component analysis (ICA) for data reduction, followed by k-means clustering and agglomerative hierarchical clustering (AHC) for unsupervised learning.
• Utilized LE and clustering for tumor grade assessment and MRSI data segmentation.

Main Results:

• LE combined with unsupervised clustering achieved 93% accuracy in classifying glioma grades II and IV, and 100% accuracy in distinguishing tumor from normal spectra.
• LE demonstrated superior data distribution linearity and cluster stability for (1) H MRSI data compared to ICA.
• LE with k-means or AHC yielded 91% accuracy for tumor grading and 100% for normal tissue identification, enabling color-coded visualization of brain and tumor regions.

Conclusions:

• Laplacian eigenmaps (LE) show significant promise for unsupervised clustering of (1) H MRS and (1) H MRSI data.
• The LE method facilitates accurate separation of brain and tumor tissues and automated color-coded visualization.
• LE offers an improved approach for analyzing brain tumor spectroscopy data, aiding in diagnosis and surgical planning.