Diffusion modelling of polarimetric cervical tissue properties

Within machine learning, diffusion offers a promising approach for modelling underlying data distributions. In this work, we use and extend neural diffusion processes to learn spatial relationships in decomposed polarimetry features, e.g., depolarisation, capturing how these properties vary across tissue samples. By conditioning on spatial coordinates and observed decomposed polarimetry values in other regions of a tissue sample, our model infers missing or uncertain regions, analogous to an infilling task. Our approach provides a principled way to assess whether local measurements align with expected spatial patterns. This enables both anomaly detection and data repair in polarimetric imaging, providing a flexible, uncertainty-aware method for interpreting and improving polarimetric measurements.