A Polarization-Imaging-Based Machine Learning Framework for Quantitative Pathological Diagnosis of Cervical Precancerous Lesions

Abstract

Polarization images encode high resolution microstructural information even at low resolution. We propose a framework combining polarization imaging and traditional microscopy imaging, constructing a dualmodality machine learning framework that is not only accurate but also generalizable and interpretable. We demonstrate the viability of our proposed framework using the cervical intraepithelial neoplasia grading task, providing a polarimetry feature parameter to quantitatively characterize microstructural variations with lesion progression in hematoxylin-eosin-stained pathological sections of cervical precancerous tissues. By taking advantages of polarization imaging techniques and machine learning methods, the model enables interpretable and quantitative diagnosis of cervical precancerous lesion cases with improved sensitivity and accuracy in a lowresolution and wide-field system. The proposed framework applies routine image-analysis technology to identify the macro-structure and segment the target region in H&Estained pathological images, and then employs emerging polarization method to extract the micro-structure information of the target region, which intends to expand the boundary of the current image-heavy digital pathology, bringing new possibilities for quantitative medical diagnosis.

INTRPRT: A Systematic Review of and Guidelines for Designing and Validating Transparent AI in Medical Image Analysis

“…were frequently used to establish transparent systems. There existed two main ways to use human-understandable features in medical imaging: 1) Extracting hand-crafted features, e. g., morphological and radiomic features, from predicted segmentation masks generated by a nontransparent model [26,27,33,61,63,47,86,50,78,104] followed by analysis of those hand-crafted features using a separate classification module; 2) Directly predicting human-understandable features together with the main classification and detection task [55,44,45,77,100]. In these approaches, all tasks usually shared the same network architecture and parameter weights.…”

Angular-Based Mueller Matrix Polarimetry Parameters for Subwavelength Pore Size Differentiation

“…When light interacts with samples, the polarization state of light may change due to scattering, absorption, refraction, and other optical phenomena; such changes in the polarization state before and after light interaction can be comprehensively described using the Mueller matrix. Scholars have exploited Mueller matrix polarimetry to analyze various materials and biological samples because the Mueller matrix encodes rich microstructure information [5][6][7][8]. Existing studies prove that Mueller matrix polarimetry can differentiate cancerous tissues [6,7], liver fibrosis [9], selected species of algae [10], and aerosol particles [11].…”

“…Scholars have exploited Mueller matrix polarimetry to analyze various materials and biological samples because the Mueller matrix encodes rich microstructure information [5][6][7][8]. Existing studies prove that Mueller matrix polarimetry can differentiate cancerous tissues [6,7], liver fibrosis [9], selected species of algae [10], and aerosol particles [11].…”

“…PBPs are then linearly combined to create polarization feature parameters (PFPs), which are much more microstructure-specific. This approach is proven useful in pathological samples [4][5][6]. To differentiate more complex samples, nonlinear models in machine learning could be utilized, but such models are often uninterpretable, and the results are not generalizable.…”

“…Given the sample data, machine learning algorithms build a model to make classification predictions or clustering decisions. Studies have shown that by utilizing machine learning, Mueller polarimetry can classify ex vivo colon cancer, hematoxylin and eosin (H&E)stained and unstained breast cancer, H&E-stained cervical cancer, and skin cancer [5,6,31,32]. Using deeper models and convolution layers that specialize on processing imagery data, the models trained using polarimetric data are capable of transforming between polarimetric imaging and brightfield imaging, synthesizing polarization-sensitive optical coherence tomography images from OCT images and classifying objects in degraded environments [33][34][35].…”