Reconstruction Accuracy vs. Discriminative Power: Spectral Unmixing Performance in Brain Tissue Hyperspectral Imaging

Area of Science:

• Medical imaging
• Biophotonics
• Computational pathology

Background:

• Hyperspectral imaging (HSI) offers potential for real-time brain tissue analysis during surgery.
• High data dimensionality of HSI hinders its clinical application.
• Spectral unmixing can reduce HSI data complexity into interpretable maps.

Purpose of the Study:

• To evaluate spectral unmixing models and normalization strategies for intraoperative brain tissue characterization.
• To assess the impact of data compression on tissue classification accuracy.
• To identify optimal configurations for clinical deployment of HSI in neurosurgery.

Main Methods:

• Four spectral unmixing models (ASM, LMM, PPNMM, FBM) and three normalization strategies (SNV, L2, Max) were tested.
• Two in vivo hyperspectral brain surgery datasets (HELICoiD, SLIMBRAIN) were utilized.
• Support Vector Machines (SVMs) were trained on abundance vectors and full spectra to evaluate spectral reconstruction (SAM) and tissue classification (F1-score).

Main Results:

• SNV normalization consistently improved classification performance across models.
• The Absorption and Scattering Model (ASM) demonstrated significant spectral reconstruction improvements over linear models.
• ASM-SNV achieved comparable classification accuracy to full spectra with substantial data compression.
• Task-agnostic reconstruction (ASM-Max) negatively impacted tumor classification, highlighting the need for task-specific optimization.

Conclusions:

• The ASM-SNV configuration provides a robust balance between data interpretability and classification accuracy for intraoperative guidance.
• Superior spectral reconstruction fidelity does not always translate to improved diagnostic discrimination.
• Task-optimized unmixing frameworks are crucial for addressing spectral variability in pathologies like brain tumors.