Application of Polarization-Imaging-Based Unsupervised Learning Method in Differential Diagnosis of Intrahepatic Cholangiocarcinoma and Colorectal Cancer Liver Metastasis

Differentiating intrahepatic cholangiocarcinoma (ICC) from colorectal cancer liver metastases (CRLM) is challenging, as current diagnostic methods rely on extensive immunohistochemistry and costly sequencing. Polarization imaging offers a cost-effective, non-invasive alternative, providing detailed morphological information through its Mueller matrix. This study included 40 ICC and 40 CRLM samples, encompassing 400 regions of interest (ROIs), which were divided into Cohort A and B for validation using Mueller matrix microscopes. A polarization super-pixel strategy applying the minibatch K-Means algorithm was used to compress the volume of polarimetric data while preserving its main polarization features. Polarization basis parameters (PBP) were extracted, followed by Uniform Manifold Approximation and Projection (UMAP) to cluster polarization pixels into subtypes. UMAP identified 6 polarization feature clusters in Cohort A and 7 in Cohort B, correlating with histological structures such as cell nucle, cytoplasm, and collagen fibers. The nuclear-associated clusters exhibited the highest classification performance, with area under the receiver operating characteristic curve (AUROC) values of 80.70% (Cluster A1 in Cohort A) and 80.10% (Cluster B1 in Cohort B). The nuclear-associated polarization date, as a polarization feature marker, demonstrated excellent discriminatory power, making this approach a promising tool for distinguishing ICC from CRLM using Mueller microscopy and unsupervised learning (UL).