VITAL: Value-Invariant Transformation and Alignment Learning for quantitative photoacoustic microscopy

Fast optical-resolution photoacoustic microscopy (OR-PAM) enables dynamic microvascular imaging but suffers from spatial misalignment and biased functional quantification that are inadequately solved by conventional registration methods. We propose VITAL (Value-Invariant Transformation and Alignment Learning), a cascaded PA signal amplitude-preserving (PSAP) registration framework that separates geometric alignment from preservation of discrete PA signal-amplitude distributions. VITAL combines interpretable keypoint-based global alignment using SuperPoint, LightGlue, and thin-plate spline transformation with topology-preserving dense refinement for sub-pixel accuracy. To reduce interpolation-induced signal distortion, nearest-neighbor warping is adopted to avoid interpolation-generated PA signal amplitudes. Validated on in vivo mouse brain imaging under oxygen challenge and controlled synthetic deformation, fine-tuned VITAL achieves an 89.1% reduction in mean squared error (MSE), improves normalized cross-correlation (NCC) from 0.178 to 0.821, and maintains a PA signal-amplitude distribution fidelity rate (PFr) of 97.83%, supporting robust downstream functional OR-PAM analysis.