A polarization super-pixel framework for extracting polarization features from Mueller matrix images

Mueller matrix describes the polarization properties of the samples comprehensively, characterizing microstructural information at subcellular-level. Mueller matrix microscopy is a promising non-invasive tool for pathological diagnosis, but it can be challenging to extract polarization parameters that correlate with pathological variation. In this study, we propose a polarization super-pixel based polarization feature extraction framework. Polarization super-pixels are able to represent the polarization features of the biological sample in a simple, compact, and comprehensive way, while reducing the data volume drastically. Using various pathological samples including breast cancer, liver cancer, and lung cancer, we show that polarization super-pixel approach greatly increases the efficiency and performance of the downstream supervised and unsupervised learning tasks, for cancerous tissue identification and microstructural composition analysis. We also propose the super-pixel based label spreading method, which iteratively propagates pathologist's initial manual label of cancerous region to the entire field of view, highlighting the tissues with the same microstructural features.