Retardance2Retardance: unsupervised denoising for polarization-sensitive optical coherence tomography

Polarization-sensitive optical coherence tomography (PS-OCT) enables advanced micron-scale characterization of biological tissues. However, its quantitative diagnostic utility is fundamentally limited by the presence of speckle and associated polarimetric noise. Conventional spatial filtering techniques suppress speckle effects yet compromise spatial resolution and can introduce severe coordinate-transformation artifacts in catheter-based applications. To address these challenges, we present the Retardance2Retardance (R2R) network, a physics-constrained, self-supervised deep learning framework for robust polarimetric speckle suppression. R2R exploits spectral binning to obtain independent speckle realizations from existing data and train the model in a Noise2Noise fashion without ground truth. A physics-informed regularization in the loss function enforces that the denoised polarization effects correspond to retardance. Additionally, a physics-informed data augmentation (PIDA) strategy prevents the network from overfitting to the static polarimetric system effects inherent to the training data. Applied to independent human coronary artery pullbacks acquired in vivo, R2R effectively suppresses speckle-induced polarimetric noise while preserving fine polarimetric features. Furthermore, the model demonstrates robust generalization to unseen complex intravascular structures, including bifurcations and deployed stents. Our R2R framework establishes a highly adaptable approach for high-fidelity, quantitative PS-OCT, including intravascular polarimetry, through effective polarimetry noise suppression and better preservation of spatial resolution.