Denoising broadband CARS hyperspectral images on variance-stabilized data: a unified benchmark

Abstract

Broadband coherent anti-Stokes Raman scattering (BCARS) enables label-free hyperspectral chemical imaging, but weak signal components are often obscured by heteroscedastic noise. Here, we develop a physics-based variance stabilization transform and benchmark six hyperspectral denoising methods to identify effective strategies for BCARS noise reduction. The transform yields nearly intensity-independent noise variance while preserving the dispersive Raman response. Within this framework, we compare singular value decomposition, BM4D, Noise2Noise, deep image prior, self-supervised diffusion, and a cross-covariance-based singular value decomposition method (CCV) introduced here, using a shared phase-retrieval pipeline. Evaluation includes five simulated hyperspectral cubes with exact ground truth and experimental measurements of glycerol, polystyrene beads, and Caenorhabditis elegans. In simulations, variance stabilization reduces the ratio of maximum to minimum channel noise variance from approximately 72 to 1.2. Deep image prior and self-supervised diffusion achieve the highest mean foreground peak signal-to-noise ratios. CCV matches the mean foreground structural similarity of deep image prior with substantially shorter runtime and no per-cube rank tuning. Experimental comparisons reveal sample-dependent performance. These results establish physics-based variance stabilization as a common framework for adapting and evaluating hyperspectral denoisers for BCARS, with CCV offering a practical balance of reconstruction fidelity, computational speed, and automated operation.