Extraction of artefactual MRS patterns from a large database using non-negative matrix factorization.

Area of Science:

• Neuroimaging
• Biomedical Engineering
• Machine Learning

Background:

• Automated pattern recognition is successful in brain tumor diagnosis using Magnetic Resonance Spectroscopy (MRS).
• Automated data quality assessment for MRS in neuro-oncology remains underdeveloped, with human expert assessment being the current standard.
• MRS data artefacts can negatively impact pattern recognition accuracy and escape human detection.

Purpose of the Study:

• To develop an automated method for detecting artefacts in MRS data for neuro-oncology.
• To enhance the reliability of tumor characterization and the robustness of pattern recognition classifiers.
• To establish more trustworthy MRS data processing and analysis pipelines.

Main Methods:

• Application of Convex Non-Negative Matrix Factorization (a variant of blind source separation) in an unsupervised manner.
• Utilizing feature extraction to distinguish between good and bad quality spectra.
• Analysis of a large database (>2000 spectra) of short and long echo time 1H-MRS data from various tumor types and anomalous masses at 1.5T.

Main Results:

• The study provides a proof of concept for using Convex Non-Negative Matrix Factorization to identify signal sources within MRS data.
• It is hypothesized that some of these identified sources correspond to signal artefacts, while others represent distinct tumor patterns.
• This approach offers a potential solution for automated artefact detection, complementing existing methods.

Conclusions:

• The proposed unsupervised method using Convex Non-Negative Matrix Factorization shows promise for automated MRS data quality assessment in neuro-oncology.
• Successful implementation could lead to more reliable tumor classifications and robust machine learning models.
• This technique has the potential to significantly improve the analysis of MRS data in clinical practice.