Artery segmentation and atherosclerotic plaque quantification using AI for murine whole slide images stained with oil red O

Fig. 1. (a) Artery and plaque segmentation combines supervised and unsupervised learning methods for automatic lesion characterization. Initially, input WSIs are inserted into a segmentation network, which was trained in a supervised manner on a large hand-labeled dataset. This yields a binary segmentation mask of the artery area. The (cropped) arterial region is then processed by the W-Net, an unsupervised network trained on a selected dataset for analysis purposes. After post-processing, both the artery and plaque regions can be used for statistical analysis. (b) Individual steps for artery and plaque segmentation in detail. The upper and middle panel illustrates the region of interest (ROI) pipeline and the ensemble structure for artery segmentation. The lower panel illustrates the proposed strategy for plaque segmentation.

Fig. 2. Qualitative examples of the individual network, the ROI pipeline, as well as the ensemble with different \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha$$\end{document} overlapping parameters. TP = white, FP = orange, FN = blue, TN = black. GT = white + blue, prediction = white + orange. The mIoU is depicted below each sample along with a color bar. The green box in the input image shows the region of interest (ROI) used by the ROI pipeline.

Fig. 3. Examples showing plaque contents (detected plaque shown in orange, arterial region in white) according to hand-labeled reference (HLR), W-Net output mask, and W-Net output mask after post-processing. The plaque amount is depicted below each sample as a percentage of the artery mask.

Fig. 4. Linear regression of predicted plaque sizes (in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hbox {mm}^2$$\end{document}) by the W-Net, with and without post-processing and color transfer, compared to the threshold-based reference (TBR) per mouse under chow or western type diet (WTD) from published reference data9. The Pearson correlation coefficient r is depicted in the top left corner. Black lines resemble perfect linear regression.

Fig. 5. Comparison of our AI-based segmentation pipeline with published reference data9. The data points represent the medians per mouse for lesion areas of individual WSIs. The student’s t-test was used to calculate p-values between groups. n.s. not significant, WTD western type diet, chow control diet.