A novel semi-supervised methodology for extracting tumor type-specific MRS sources in human brain data

Area of Science:

• Biomedical Engineering
• Medical Imaging Analysis
• Computational Biology

Background:

• Clinical investigation of human brain tumors relies on imaging for extent and location but lacks biochemical insights.
• Magnetic Resonance Spectroscopy (MRS) offers a metabolic fingerprint of tissue, complementing standard imaging.
• Analyzing single-voxel MRS data presents challenges in signal source identification due to tissue heterogeneity.

Purpose of the Study:

• To develop and evaluate a novel semi-supervised approach for source identification in single-voxel magnetic resonance spectra of brain tumors.
• To improve the accuracy of brain tumor classification by leveraging prior class information within a matrix factorization framework.

Main Methods:

• Utilized a convex variant of non-negative matrix factorization (NMF) capable of handling negative data values.
• Proposed a novel semi-supervised NMF approach integrating prior class knowledge into model optimization.
• Integrated class-specific information by defining metrics within a latent variable space for matrix factorization.

Main Results:

• The proposed semi-supervised method demonstrated superior performance compared to unsupervised NMF.
• Achieved near-perfect correlation between extracted sources and mean spectra of tumor types.
• Showcased improved tissue type classification accuracy for brain tumors.

Conclusions:

• Semi-supervised learning, by integrating class information into unsupervised matrix factorization, significantly benefits discriminative source identification.
• The methodology enhances brain tumor labeling from single-voxel spectroscopy data.
• The proposed approach holds potential for broader applications in biomedical signal processing.