ISSN 1000-1239 CN 11-1777/TP

计算机研究与发展 ›› 2015, Vol. 52 ›› Issue (9): 2135-2144.doi: 10.7544/issn1000-1239.2015.20140683

• 人工智能 • 上一篇    下一篇



  1. (广东工业大学自动化学院 广州 510006) (
  • 出版日期: 2015-09-01
  • 基金资助: 

A Machine Learning Method for Histopathological Image Automatic Annotation

Zhang Gang, Zhong Ling, Huang Yonghui   

  1. (School of Automation, Guangdong University of Technology, Guangzhou 510006)
  • Online: 2015-09-01

摘要: 病理图像能够揭示疾病的原因及严重程度,在临床诊断中有重要应用.病理图像中局部区域与病理特性之间不明确的对应关系为建立计算机辅助诊断模型带来了困难.基于全局图像特征表达和等分小块等方法难以有效表达病理特性的局部性.提出一种基于多示例多标签学习的活检病理图像自动标注框架,对病理特性的局部性进行表达.通过带区域约束条件的分割算法把病理图像划分为若干视觉上不连续的区域,对区域进行基于纹理和内部结构的特征提取,把病理图像转化为多示例样本,在此基础上提出一种基于贝叶斯学习的多示例多标签稀疏集成算法.在本地大型三甲医院的皮肤科活检样本数据集上进行方法有效性评估,结果表明该方法能得到医学上可接受的标注准确率,从而说明其有效性.

关键词: 病理图像标注, 病理特性识别, 多示例多标签学习, 标准分割, 图像特征提取

Abstract: Histopathological image can reveal the reason and severity of diseases, which is important for clinical diagnosis. Automatic analysis of histopathological image may release doctor’s burden for manual annotation which can preserve more time for doctors to focus on special and difficult cases. However, the ambiguous relationship between local regions in a histopathological image and histopathological characteristics makes it difficult to construct a computer-aid model. An automatic annotation method for histopathological images based on multiple-instance multiple-label (MIML) learning is proposed, aiming at directly modeling the medical experience of doctors, which suggests that each annotated term associated with an image corresponds to a local visually recognized region. We propose a self-adaptive region cutting method with constraints, to segment each image into several visually disjoint regions, and then perform a feature extraction for each generated region based on texture and inner structures. The whole image is regarded as a bag and regions as instances, thus an image is expressed as a multiple-instance sample. Then we propose a sparse ensemble multiple-instance multiple-label learning algorithm, S-MIMLGP, based on Bayesian learning, and compare it with current multiple-instance single label and multiple-instance multiple-label algorithms. The evaluation on a clinical dataset from the dermatology of a large local hospital shows that the proposed method can yield medically acceptable annotation accuracy, hence indicates its effectiveness.

Key words: histopathological image annotation, histopathological characteristics recognition, multiple-instance multiple-label learning, normalized cut, image feature extraction