logo

SCIENCE CHINA Information Sciences, Volume 60, Issue 9: 092108(2017) https://doi.org/10.1007/s11432-016-9018-7

High-throughput fat quantifications of hematoxylin-eosin stained liver histopathological images based on pixel-wise clustering

More info
  • ReceivedMar 21, 2016
  • AcceptedJan 16, 2017
  • PublishedApr 28, 2017

Abstract

Besides diagnosis of fatty liver disease (FLD) using multiple medical imaging techniques in clinic, accurate fat quantification of liver tissue slice, especially the fat droplets measurement, is still a critical indicator in related pathological researches. Stained by hematoxylin-eosin (HE), different tissue components with different colors need to be identified and measured manually in conventional approaches. Automated liver fat quantification of HE stained images remains challenging because forms and distributions of fat are extremely irregular with no clear boundaries, especially in conducting high-throughput analysis which demands quick processing and higher accuracy for the reference of pathologists. To solve this problem, we propose an automated liver fat quantifications pipeline of HE stained images based on pixel-wise clustering, which firstly extracts high-relevant pixel-level features with color mode transformation, then locates boundaries between nuclei, fat and other components by clustering image pixels in an unsupervised mode, and finally identifies indicative fat droplets based on a set of morphological criteria. The pipeline was verified in analysis of multifold fatty liver treatment assays, with experimental results showing high accuracy and adaptability in fat droplets quantification despite data variance. Quantitative indicators provide a reliable evidence for relevant pathological researches or therapy selection, in which number and average area of indicative fat droplets increased sharply in severe and moderate-grade FLD respectively. Those indicators might be utilized as surrogate biomarkers for further researches.


Acknowledgment

This work was supported by National Natural Science Foundation of China (Grant No. 61501121), Scientific Research Foundation for the Returned Overseas Chinese Scholars, State Education Ministry (Grant No. (2015)1098), Provincial Science Foundation, Fujian Provincial Department of Science and Technology (Grant No. 2015J05145), and Provincial Research Funds for Innovative Youth, Fujian Provincial Department of Education (Grant No. JA14084).

  • Figure 1

    (Color online) Original HE stained images of liver hispathological slices. (a) Normal liver; (b) low-grade; protectłinebreak (c) moderate-grade; and (d) severe-grade fatty liver.

  • Figure 2

    Workflow of proposed automated liver fat quantification pipeline.

  • Figure 3

    (Color online) Illustration of color mode transformation. (a) An original HE image of a normal liver tissue slice, and its decomposed signal intensity map after color mode transformation, where R, G and B signals shown from top to bottom in the RGB column, and L, a and b signals shown from top to bottom in the Lab column; (b) a severe-grade fatty liver tissue slice HE image, and its RGB and Lab components respectively.

  • Figure 4

    (Color online) Illustrations of HE stained image segmentation based on k-means clustering. (a) An original HE image of a normal liver tissue slice; (b) distribution map of all pixels before clustering in the feature space as cyan, where the polar system represent as $L$ and $b$ values distributions in the coordinates, in which the radial coordinate is the root mean square of $L$ and $b$, and the angular coordinate is the intersection angle between $L$ and $b$ values; (c) distribution map of pixels after clustering, where dark gray dots represent as pixels of fat, light gray dots as cytoplasm and black dots as nuclei; protectłinebreak (d) color map of labeled liver image pixels after k-means clustering; (e)–(h) corresponding maps of a severe-grade fatty liver tissue slice HE image.

  • Figure 5

    (Color online) Steps of indicative fat droplets identification. (a) A zoom-in original image including touching fat areas; (b) grey-scale intensity map of labeled fat areas; (c) highlighted segmented liver fat droplet by Wastershed; (d) histogram of grey-scale intensity map; (e) candidate fat droplets using adaptive global grey-scale threshold; protectłinebreak (f) indicative fat droplets after filtering with shape model marked as orange on the original image.

  • Figure 6

    (Color online) Results of nuclei and fat segmentation on routine HE stained liver histopathologic images (a)–(d), where the original HE images, the labeled pixels after clustering, the segmented nuclei and fat areas, and the zoom in segmentation results are shown in each column respectively. Boundaries of nuclei and fat are marked as blue and red lines respectively in segmentation results. (e) Histograms and error bars of nuclei and fat area proportions, where $N$ represents as nuclei and $F$ as fat. (f) Variation tendency of nuclei and fat area proportions calculated based on image segmentation.

  • Figure 7

    (Color online) Relationships between FLD severity and nuclei and fat distributions. (a) Indicative fat droplet number and (b) average area of indicative fat droplets manually (M) and automatically (A) detected from left to right.

Copyright 2020 Science China Press Co., Ltd. 《中国科学》杂志社有限责任公司 版权所有

京ICP备18024590号-1