# Unsupervised learning of metabolic fingerprints from 3D magnetic resonance spectroscopic imaging enables glioma subtype classification

## TL;DR

This study shows that non-invasive 3D MRI combined with machine learning can accurately classify brain tumor subtypes based on their metabolic profiles.

## Contribution

A novel unsupervised learning framework using G-NMU and UMAP for non-invasive glioma classification based on metabolic fingerprints.

## Key findings

- The framework achieved 99.65% accuracy and 99.07% AUC in classifying glioma subtypes.
- Metabolic fingerprints of 2HG, serine, and inositol were identified as key drivers for classification.
- G-NMU outperformed traditional spectral fitting by not requiring prior assumptions about tumor metabolism.

## Abstract

Accurate classification of glioma subtypes is essential for personalized treatment, yet current diagnostic approaches rely on invasive procedures to determine molecular profiles. This study aims to enhance non-invasive glioma classification by integrating metabolic imaging with advanced unsupervised learning.

Whole-brain 3D Magnetic Resonance Spectroscopic Imaging (MRSI) was performed at 3 Tesla. From 26 scanned patients, 12 gliomas (5 astrocytomas, 5 oligodendrogliomas, 2 glioblastomas) that passed strict quality-control criteria were included for analysis. Spectral decomposition was performed using Global Non-Negative Matrix Underapproximation (G-NMU), and tumor subtype classification was achieved with Uniform Manifold Approximation and Projection (UMAP) followed by K-means clustering.

The proposed framework was able to classify tumor types with an accuracy of 99.65% and an AUC of 99.07. Clear subtype-specific metabolic fingerprints were validated by hierarchical clustering and UMAP embeddings, emphasizing 2HG, serine, and inositol as important classification drivers.

This study demonstrates that whole-brain MRSI spectral decomposition based on G-NMU is a reliable non-invasive method for classifying gliomas. In contrast to spectral fitting on prior-knowledge basis sets, G-NMU accurately separates astrocytoma, oligodendroglioma, and glioblastoma by extracting metabolic features without making assumptions about the tumor metabolic composition. These results suggest that integration of metabolic imaging and unsupervised learning into clinical workflows may improve molecular stratification for noninvasive glioma diagnosis.