Polarization memory mapping using physically informed neural network

We present a machine learning approach to cancer tissue detection based on physically meaningful features derived from polarized light interactions. Instead of relying on image-based segmentation with labeled data, we use unsupervised models trained on pixel-wise maps of Stokes vectors, phase differences, and polarization metrics. These features, extracted from polarization-resolved data at varying tissue depths, enable clustering and anomaly detection without histopathological ground truth. This reduces reliance on time-consuming labeling by pathologists. Our approach highlights the diagnostic potential of polarization-based signatures and shows a path toward more interpretable, label-free AI methods in biomedical imaging.