欢迎来到《四川大学学报(医学版)》

基于无监督学习的数字病理切片自动分割方法

秦航宇 邓杨 周燕燕 刘洪红 李丽 周琪琪 梅娟 步宏 包骥

秦航宇, 邓杨, 周燕燕, 等. 基于无监督学习的数字病理切片自动分割方法[J]. 四川大学学报(医学版), 2021, 52(5): 813-818. doi: 10.12182/20210960203
引用本文: 秦航宇, 邓杨, 周燕燕, 等. 基于无监督学习的数字病理切片自动分割方法[J]. 四川大学学报(医学版), 2021, 52(5): 813-818. doi: 10.12182/20210960203
QIN Hang-yu, DENG Yang, ZHOU Yan-yan, et al. Automatic Segmentation of Digital Pathology Slides Based on Unsupervised Learning[J]. JOURNAL OF SICHUAN UNIVERSITY (MEDICAL SCIENCE EDITION), 2021, 52(5): 813-818. doi: 10.12182/20210960203
Citation: QIN Hang-yu, DENG Yang, ZHOU Yan-yan, et al. Automatic Segmentation of Digital Pathology Slides Based on Unsupervised Learning[J]. JOURNAL OF SICHUAN UNIVERSITY (MEDICAL SCIENCE EDITION), 2021, 52(5): 813-818. doi: 10.12182/20210960203

基于无监督学习的数字病理切片自动分割方法

doi: 10.12182/20210960203
基金项目: 科技部重点研发计划重点专项(No. 2017YFC0113908),北京精鉴病理发展基金会(No. 2019-0007),四川省国际科技合作与交流研发项目(No. 2017HH0070、No.2018HH0037),成都市新型产业技术研究院技术创新项目(No. 2017-CY02-00026-GX),四川大学华西医院临床研究孵化项目(No. 2020HXFH029)和四川大学华西医院学科卓越发展1·3·5 工程项目(No. ZYGD18012)资助
详细信息
    通讯作者:

    E-mail:baoji@scu.edu.cn

Automatic Segmentation of Digital Pathology Slides Based on Unsupervised Learning

More Information
  • 摘要:   目的  使用无监督的方式进行图像分割,作为人工标记的一种替代。  方法  选取了共100张HE染色和巴氏染色切片的全片数字化图像(whole slide image, WSI)数据作为研究和测试的对象,其中乳腺切片70张,肺切片20张,甲状腺切片10张。为了保证数据的多样性,乳腺的切片包含了正常组织、炎症、肿瘤,肺切片取材主要为下叶新生物(包含了炎症和肿瘤),甲状腺为细针穿刺的细胞(均为良性)。每张图像的最大总倍率(原始倍率)均为400倍,文件格式为ndpi。对每张WSI进行人工的标注,每张WSI的标注区域都大于10个视野,标注后的信息将用于有效性的验证。使用基于超像素与全卷积神经网络的算法构建无监督图像分割技术,对没有标记的WSI的任意感兴趣区域(regions of interest, ROI)进行图像分割。与区域邻接图合并的方法进行比较,以欠分割错误差率、边缘召回率和平均交并结果比判定两种方法的分割效果,并比较两种方法的效率。在执行效率的比较中,测试过程包含了超像素的预处理的时间,去掉了加载深度学习引擎的时间。  结果  对WSI任意ROI区域按纹理和颜色对图像实现了无监督的自动分割,乳腺切片、肺切片和甲状腺切片测试的结果差异小,多次测试的结果稳定,但该方法在对炎症和肿瘤的区分中表现一般。其欠分割错误差率、边缘召回率和平均交并结果分别为19.10%、82.06%和45.06%。区域邻接图合并的方法的欠分割错误差率、边缘召回率和平均交并的结果分别为21.52%、78.39%和44.81%。在GPU模式下整个过程平均耗时为0.27 s,在CPU模式下平均耗时为1.30 s,由于区域邻接图合并的方法没有实现GPU模式,在CPU模式下平均耗时为10.5 s。  结论  本方法通过简单的人机交互操作得到理想的像素级标注结果,可以有效降低数字病理切片数据标注的成本,比区域邻接图合并的方法在处理图像纹理的方面表现得更好,处理速度更快。
  • 图  1  用于训练的卷积神经网络模型

    Figure  1.  Convolutional neural network model used for training

    图  2  超像素分割、合并、选取的实验图组

    Figure  2.  Experimental group of superpixel segmentation, merging and selection

    A is the original image, from left to right are images with three magnification rates of 12.5, 200, and 400 respectively; B is the superpixel obtained after calculatiion of image A, and is expressed in yellow color; C is the final segmentation result obtained based on RAG; D is the final segmentation result obtained by our proposed method. The magnification of B, C and D is 200 times on the left and 400 times on the right.

  • [1] 叶美华, 盛弘强, 王怡栋, 等. 数字病理切片系统可视化数据应用简介. 中华病理学杂志,2012,41(1): 66–68. doi: 10.3760/cma.j.issn.0529-5807.2012.01.020
    [2] LONG J, SHELHAMER E, DARRELL T. Fully convolutional networks for semantic segmentation. arXiv, 2015: 1411.4038[2020-11-03]. https://arxiv.org/abs/1411.4038v1.
    [3] 向日华, 王润生. 一种基于高斯混合模型的距离图像分割算法. 软件学报,2003,14(7): 1250–1257.
    [4] SHIRAZI M N, NODA H. A deterministic iterative algorithm for HMRF-textured image segmentation//Proceedings of 1993 International Conference on Neural Networks (IJCNN-93-Nagoya, Japan). Nagoya, Japan: IEEE, 1993: 2189-2194.
    [5] 侯彪, 徐婧, 刘凤, 等. 基于第二代Bandelet域隐马尔可夫树模型的图像分割. 自动化学报,2009,35(5): 498–504.
    [6] 霍迎秋, 秦仁波, 邢彩燕, 等. 基于CUDA的并行K-means聚类图像分割算法优化. 农业机械学报,2014,45(11): 47–53. doi: 10.6041/j.issn.1000-1298.2014.11.008
    [7] ACHANTA R, SUSSTRUNK S. Superpixels and polygons using simple non-iterative clustering//2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Honolulu, USA: IEEE, 2017: 4895–4904.
    [8] HADJI I, WILDES R P. What do we understand about convolutional networks? arXiv, 2018: 1803.08834[2020-11-03]. https://arxiv.org/abs/1803.08834.
    [9] TANG Y, ZHAO L, REN L. Different versions of entropy rate superpixel segmentation for hyperspectral image//2019 IEEE 4th International Conference on Signal and Image Processing (ICSIP). Wuxi,China: IEEE, 2019:1050-1054.
    [10] ACHANTA R, SHAJI A, SMITH K, et al. SLIC superpixels compared to state-of-the-art superpixel methods. IEEE Transactions Pattern Analysis Machine Intelligence,2012,34(11): 2274–2282. doi: 10.1109/TPAMI.2012.120
    [11] BEJNORDI B E, LITJENS G, HERMSEN M, et al. A multi-scale superpixel classification approach to the detection of regions of interest in whole slide histopathology images//Proceedings Volume 9420, Medical Imaging 2015: Digital Pathology. Orlando, USA: SPIE Medical Imaging, 2015: 94200H.
    [12] CIMPOI M, MAJI S, KOKKINOS I, et al. Deep filter banks for texture recognition, description, and segmentation. arXiv, 2015: 1507.02620[2020-11-03]. https://arxiv.org/abs/1507.02620.
    [13] KANEZAKI A. Unsupervised image segmentation by backpropagation//2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). Calgary, Canada: IEEE, 2018: 1543–1547.
    [14] GALLOWAY A, GOLUBEVA A, TANAY T, et al. Batch normalization is a cause of adversarial vulnerability. arXiv, 2019: 1905.02161[2020-11-03]. https://arxiv.org/abs/1905.02161.
    [15] 罗学刚, 吕俊瑞, 彭真明. 超像素分割及评价的最新研究进展. 激光与光电子学进展,2019,56(9): 53-63.
    [16] GHORBANZADEH O, BLASCHKE T, GHOLAMNIA K, et al. Evaluation of different machine learning methods and deep-learning convolutional neural networks for landslide detection. Remote Sens,2019,11(2): 196[2020-11-03]. https://doi.org/10.3390/rs11020196.
    [17] TREMEAU A, COLANTONI P. Regions adjacency graph applied to color image segmentation. IEEE Transactions Image Processing,2000,9(4): 735–744. doi: 10.1109/83.841950
    [18] SCHUURMANS M, BERMAN M, BLASCHKO M B. Efficient semantic image segmentation with superpixel pooling. arXiv, 2018: 1806.02705[2020-11-03]. https://arxiv.org/abs/1806.02705.
    [19] 蔡莉, 王淑婷, 刘俊晖, 等. 数据标注研究综述. 软件学报,2020,31(2): 302–320.
    [20] 薛腾飞, 傅群超, 王枞, 等. 基于区块链的医疗数据共享模型研究. 自动化学报,2017,43(9): 1555–1562.
    [21] CAMPANELLA G, HANNA M G, GENESLAW L, et al. Clinical-grade computational pathology using weakly supervised deep learning on whole slide images: 8. Nat Med,2019,25(8): 1301–1309. doi: 10.1038/s41591-019-0508-1
  • 加载中
图(2)
计量
  • 文章访问数:  95
  • HTML全文浏览量:  30
  • PDF下载量:  12
  • 被引次数: 0
出版历程
  • 收稿日期:  2020-11-03
  • 修回日期:  2021-02-05
  • 网络出版日期:  2021-12-06
  • 刊出日期:  2021-09-20

目录

    /

    返回文章
    返回