Deep Learning for Histopathological Image Analysis: Towards Computerized Diagnosis on Cancers

Automated detection and segmentation of histologic primitives are critical steps for developing computer-aided diagnosis and prognosis system on histopathological tissue specimens. For a number of cancers, the clinical cancer grading system is highly corr

  • PDF / 1,481,455 Bytes
  • 23 Pages / 439.37 x 666.142 pts Page_size
  • 102 Downloads / 316 Views

DOWNLOAD

REPORT


Deep Learning for Histopathological Image Analysis: Towards Computerized Diagnosis on Cancers Jun Xu, Chao Zhou, Bing Lang and Qingshan Liu

Abstract Automated detection and segmentation of histologic primitives are critical steps for developing computer-aided diagnosis and prognosis system on histopathological tissue specimens. For a number of cancers, the clinical cancer grading system is highly correlated with the pathomic features of histologic primitives that appreciated from histopathological images. However, automated detection and segmentation of histologic primitives is pretty challenged because of the complicity and high density of histologic data. Therefore, there is a high demand for developing intelligent and computational image analysis tools for digital pathology images. Recently there have been interests in the application of “Deep Learning” strategies for classification and analysis of big image data. Histopathology, given its size and complexity, represents an excellent use case for application of deep learning strategies. In this chapter, we present deep learning based approaches for two challenged tasks in histological image analysis: (1) Automated nuclear atypia scoring (NAS) on breast histopathology. We present a Multi-Resolution Convolutional Network (MR-CN) with Plurality Voting (MR-CN-PV) model for automated NAS. MR-CN-PV consists of three Single-Resolution Convolutional Network (SR-CN) with Majority Voting (SR-CN-MV) model for getting independent NAS. MR-CN-PV combines three scores via plurality voting for getting final score. (2) Epithelial (EP) and stromal (ST) tissues discrimination. The work utilized a pixel-wise Convolutional Network (CN-PI) based segmentation model for automated EP and ST tissues discrimination. We present experiments on two challenged datasets. For automated NAS, the MR-CN-PV model was evaluated on MITOS-ATYPIA-14 Challenge dataset. MRCN-PV model got 67 score which was placed the second comparing with the scores of other five teams. The proposed CN-PI model outperformed patch-wise CN (CN-PA) models in discriminating EP and ST tissues on a breast histological images.

J. Xu (B) · C. Zhou · B. Lang · Q. Liu Jiangsu Key Laboratory of Big Data Analysis Technique, Nanjing University of Information Science and Technology, Nanjing 210044, China e-mail: [email protected] © Springer International Publishing Switzerland 2017 L. Lu et al. (eds.), Deep Learning and Convolutional Neural Networks for Medical Image Computing, Advances in Computer Vision and Pattern Recognition, DOI 10.1007/978-3-319-42999-1_6

73

74

J. Xu et al.

6.1 Introduction Cancer is the leading cause of death in the United States [1] and China [2]. In 2015 alone there were 4.3 million new cancer cases and more than 2.8 million cancer deaths in China [2]. Fortunately, most of the cancers have a very high chance of cure if detected early and treated adequately. Therefore, earlier diagnosis on cancers and better prognostic prediction of disease aggressiveness and patient outcome are pretty important. The pat