Digital Pathology Based on Fully Polarized Microscopic Imaging

Yue Yao; Haojie Pei; Hao Li; Jiachen Wan; Lili Tao; Hui Ma

doi:10.3788/CJL231462

Journals >Chinese Journal of Lasers >Volume 51 >Issue 9 >Page 0907008 > Article

Chinese Journal of Lasers
Vol. 51, Issue 9, 0907008 (2024)

Digital Pathology Based on Fully Polarized Microscopic Imaging

Yue Yao^1、2, Haojie Pei^1、2, Hao Li³, Jiachen Wan^1、2, Lili Tao³, and Hui Ma^1、2、*

Author Affiliations

¹Shenzhen International Graduate School, Tsinghua University, Shenzhen 518055, Guangdong, China

²Guangdong Engineering Center of Polarization Imaging and Sensing Technology, Shenzhen 518055, Guangdong, China

³Department of Pathology, Peking University Shenzhen Hospital, Shenzhen 518036, Guangdong, China

show less

DOI: 10.3788/CJL231462 Cite this Article Set citation alerts

Yue Yao, Haojie Pei, Hao Li, Jiachen Wan, Lili Tao, Hui Ma. Digital Pathology Based on Fully Polarized Microscopic Imaging[J]. Chinese Journal of Lasers, 2024, 51(9): 0907008 Copy Citation Text

show less

Configuration of two types of transmission Mueller matrix microscopes[12]. (a) Dual rotating retarders-based Mueller matrix microscope; (b) dual division of focal plane polarimeters-based Mueller matrix microscope

Fig. 1. Configuration of two types of transmission Mueller matrix microscopes^[12]. (a) Dual rotating retarders-based Mueller matrix microscope; (b) dual division of focal plane polarimeters-based Mueller matrix microscope

Download full size

Images of polarization parameters of breast cancer tissues[38]. (a) Parameters DL, t1, Δ, D, and b correspond to the cell nuclei (labeled by the black solid line in the H&E stained image); (b) parameters rL, qL, δ, PL,and θ correspond to the fiber tissue (marked by the red solid line in the H&E stained image)

Fig. 2. Images of polarization parameters of breast cancer tissues^[38]. (a) Parameters D_L, t₁, Δ, D, and b correspond to the cell nuclei (labeled by the black solid line in the H&E stained image); (b) parameters r_L, q_L, δ, P_L,_{and θ correspond to the fiber tissue (marked by the red solid line in the H&E stained image)}

Download full size

Fig. 3. Dual-modality machine learning framework for pixel level polarization feature extraction of cervical precancerous tissues^[47]. (a) Mueller matrix light intensity image m11 and H&E image registration; (b) U-Net segmentation of cervical epithelial region in H&E image to generate mask M; (c) mapping M to PBPs to select target pixels; (d) derive a PFP by inputting target pixels in to a statistical distance-based machine learning model; (e) using PFP to distinguish normal and 3 stages of cervical precancerous lesions

Download full size

Fig. 4. Polarization parameter images based radiomics approach^[43]

Download full size

Fig. 5. Pixel clustering of Mueller matrix images from 222 ROIs of liver cancer tissues^[52]

Download full size

Fig. 6. Identification of polarization markers that correlates with hepatocellular carcinoma (HCC) differentiation degree^[52]. (a) Density heatmap of normal and malignant HCC, zoomed in on two potential polarization markers, cluster 2 and 6; (b) proportion of pixels belongs to cluster 2 in each ROI are calculated for each differentiation degree; (c) proportion of pixels belongs to cluster 6 in each ROI are calculated for each differentiation degree; (d) density heatmap for each differentiation degree, the dashed line indicates the selected polarization marker in the cytoplasm cluster to differentiate well and moderately differentiated samples; (e) in well and moderately differentiated samples, calculate the pixel proportion of the selected cytoplasm cluster

Download full size

Fig. 7. Polarization super-pixels based label extension process for lung cancer ROIs. (a) Pathologist provides a small initial label within the red box; (b) obtain extended label for the whole red box by selecting super-pixels with high contribution, pathologist identify and retain the correct label; (c) expand the red box, pathologist repeat polarization super-pixels calculation and selection to obtain extended label; (d) by iterating the above process, the entire extended label for the ROI is obtained; (e) extend the label to a new ROI 1; (f) extend the label to a new ROI 2

Download full size

Method

Applicable condition

Advantage

Disadvantage

Supervised

learning

1）Have large amount of well labeled polarization data

2）Target on clear structure recognition

1）Superior performance in recognition tasks

1）Time and labor consuming to acquire labels

2）Performance degrades with unseen data

Unsupervised

learning

1）Lack of well labeled polarization data

2）Aim to uncover underlying polarization features

1）No need for labeling

2）Can explore hidden polarization features

1）Cannot provide direct predictive results

2）Requires more computational power to handle large datasets

Polarization

super-pixel

1）Have small number of initial labels，and large amount of unlabeled data

2）Aim to uncover underlying polarization features

1）Overcome the needs for large amount of well labeled data

2）Fully utilizes polarization features to achieve fast labeling，enhances model generalization

1）May not outperform supervised learning for specific tasks

2）Performance remains constrained by the quality of initial labels

Table 1. Summary of machine learning based polarization feature extraction methods

Yue Yao, Haojie Pei, Hao Li, Jiachen Wan, Lili Tao, Hui Ma. Digital Pathology Based on Fully Polarized Microscopic Imaging[J]. Chinese Journal of Lasers, 2024, 51(9): 0907008

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information