Thesis: Research on WSIs analysis using deep learning

Research on ovarian cancer pathological image recognition technology based on deep learning

Dec 1, 2019 2:04 AM


Recent years, with the development of computer hardware technology, deep learning has made great progress in the fields of computer vision, natural language processing, machine translation, medical imaging and bioinformatics. The use of deep learning, especially supervised learning for pathological image analysis has also become a hot issue in the field of medical diagnosis. Ovarian cancer is the fourth most common cause of death among women in the world. Effective diagnosis of ovarian cancer can greatly increase the success rate of later treatment, and its pathological image-based diagnosis is regarded as the “gold standard” in most cancer diagnosis. However, the large amount of data and the complexity of the image make the diagnosis process take a long time. The serious lack of pathologists in China increases the burden on existing pathologists, which can easily cause fatigue and distraction, and affect the accuracy of diagnosis. Therefore, the research of ovarian cancer pathological image technology based on deep learning is particularly important.

Using deep learning, we can automatically extract abstract and advanced feature expressions in pathological images, and then analyze their advanced features. In addition, weakly supervised learning technology also provides a theoretical basis for using only data containing slide-level annotations. This paper aims to study the automatic analysis method of ovarian cancer pathological image based on deep learning. We build an independent database of ovarian cancer pathological images, train and optimize the model, and implement deep learning-based auxiliary diagnosis of ovarian cancer.

The main innovations and contributions of this article are summarized as follows:

  1. We constructed an independent ovarian cancer pathological image dataset with a total of 1270 whole slide images (WSIs). The dataset has a large staining span, diverse appearance and good generalization. We then describe the uniqueness of WSIs from general images, and finally describe their general preprocessing methods, including background removal, patch extraction, and multi-scale mapping.
  2. Due to the high resolution of WSIs and the limitations of computer hardware, it is necessary to extract patch images from WSIs, analyze the patch images, and then analyze WSIs indirectly. Therefore, this paper only uses WSIs and their slide-level annotations to explore the classification task of ovarian cancer pathological images and the localization task of tumors using weakly supervised learning. The experiments prove that the proposed approach in this paper has better performance in ovarian cancer pathological image analysis.
  3. In order to further analyze the differences in the tumor slides, and make our auxiliary diagnostic system more intelligent, we use unsupervised learning to analyze the tumor slides, find different clusters, analyze their differences, and prove their significance in combination with pathological knowledge.