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

Hyperspectral imaging holds promise for intraoperative brain tissue characterisation, but its high dimensionality complicates clinical deployment. Spectral unmixing offers a pathway to compress data into interpretable abundance maps, yet its impact on downstream tissue discrimination remains unclear. This study evaluated four unmixing models (ASM, LMM, PPNMM, FBM) and three normalisation strategies (SNV, L2, Max) across two hyperspectral in vivo brain surgery datasets HELICoiD (n = 15), SLIMBRAIN (n = 27). Spectral reconstruction (SAM) and tissue classification (F1-score) were assessed using optimised SVMs trained on abundance vectors versus full spectra. SNV normalisation consistently yielded superior classification performance, while the Absorption and Scattering Model (ASM) achieved median PSNR improvements exceeding 15 dB over linear models due to its scattering correction term. Notably, ASM-SNV abundance maps maintained classification accuracy comparable to full-spectrum classifiers despite an order of magnitude data compression. Conversely, superior reconstruction fidelity did not guarantee improved discrimination: ASM-Max significantly degraded tumour classification in SLIMBRAIN (p<0.001), indicating that task-agnostic reconstruction can discard clinically relevant spectral features. Tumour classification remained the most challenging task across all conditions, reflecting inherent inter-tumour spectral heterogeneity. These findings establish ASM-SNV as the best tested configuration for intraoperative guidance, balancing spectral interpretability and classification robustness while highlighting the need for task-optimised unmixing frameworks to address pathological variability.