Hierarchical non-negative matrix factorization (hNMF): a tissue pattern differentiation method for glioblastoma...

Area of Science:

• Neuroimaging
• Biomedical Engineering
• Medical Physics

Background:

• Magnetic Resonance Spectroscopic Imaging (MRSI) shows promise for glioblastoma multiforme (GBM) diagnosis and prognosis.
• Conventional non-negative matrix factorization (NMF) struggles with interpreting MRSI data with more than two tissue patterns, limiting its clinical utility.

Purpose of the Study:

• To introduce a novel hierarchical non-negative matrix factorization (hNMF) method for improved differentiation of GBM tissue patterns using MRSI data.
• To enhance the diagnostic accuracy of MRSI in GBM by addressing data interpretation challenges.

Main Methods:

• Development and application of a hierarchical non-negative matrix factorization (hNMF) algorithm.
• Blindly separating spectral sources in short-echo time (TE) proton (¹H) MRSI data through multi-level NMF, processing two tissue patterns per level.
• Validation using simulated data and in vivo short-TE ¹H MRSI data from GBM patients, with expert knowledge confirming spectral source accuracy.

Main Results:

• The hNMF method accurately estimates three distinct tissue patterns in GBM tumoral and peritumoral areas: normal, tumor, and necrosis.
• hNMF successfully overcomes the limitations of conventional NMF in complex MRSI data interpretation.
• The method generates easily interpretable maps illustrating the contribution of each tissue pattern per voxel.

Conclusions:

• hNMF provides a robust and effective approach for differentiating GBM tissue patterns from MRSI data.
• This technique offers valuable additional information for GBM diagnosis, improving upon existing methods.
• The interpretable output maps from hNMF facilitate clinical application and understanding of GBM heterogeneity.