基于深度学习方法的乳腺组织病理学图像分析研究
发布时间：2021-01-29 17:22
数字病理学是医学协议中具有挑战性的进展之一。病理检查在诊断过程中起着至关重要的作用,并使病理学家能够对微观结构进行分类。病理学家在显微镜下分析了大量的活检切片。对细胞核的组织学结构,形态变化和生物组织分布的分析有助于病理学家更好地识别组织病理学样本。高含量活检组织病理学分类和分级可提供重要的预后信息,这对于了解疾病（癌症）的扩散和预报至关重要。在主要的癌症中,乳腺癌是影响世界各地女性癌症死亡的主要原因之一,而且在发达国家和发展中国家逐渐增加。然而,在乳腺组织的全切片扫描图像（Whole slide images,WSIs）中手动诊断疾病是一项艰巨而富有挑战性的任务。为了克服手动分析的缺点,各种自动诊断系统（即,简单的图像处理算法或基于深度学习的算法）相继被开发了出来。本文主要研究开发基于深度学习的自动检测框架,这个框架能够从不同类型的组织病理学图像中对乳腺癌进行检测。世界卫生组织提出了一种乳腺癌分级标准,称为诺丁汉分级系统。它结合了三种形态学预后因素,即有丝分裂计数,肾小管形成和细胞核多形性。本文描述了有关组织病理学图像分析的理论指导,并定义了可提高组织病理学家诊断和预报能力的乳腺癌分...
【文章来源】：哈尔滨工业大学黑龙江省 211工程院校 985工程院校
【文章页数】：156 页
【学位级别】：博士
【文章目录】：
摘要
Abstract
Acronyms and symbols
Chapter 1 Introduction
1.1 Background and significances of the study
1.2 Basic theory of deep learning
1.2.1 Deep learning
1.2.2 Deep learning programming frameworks
1.3 Steps involved in biopsy slide preparation and breast cancer grading
1.3.1 Fast slide scanners for digital image acquisition
1.3.2 Standard grading system for breast cancer recognition
1.4 Research status of Histopathology image detection techniques
1.4.1 Breast cancer detection/localization methods
1.4.2 Breast cancer segmentation methods
1.4.3 Breast cancer classification methods
1.5 Motivation
1.6 Limitations and challenges in breast histology image analysis
1.7 Thesis structure
1.8 Contributions
Chapter 2 Stain color normalization of hematoxylin and eosin stained histopathologyimages
2.1 Introduction
2.2 Stain color normalization of hematoxylin and eosin stained images
2.2.1 Benchmark color normalization methods to reduce color variation in breasthistology images
2.2.2 Proposed color normalization method based on image structure and color statics
2.3 Experimental results
2.3.1 Materials
2.3.2 Metrics to check the performance of color normalization methods
2.3.3 Experimental results and performance analysis
2.4 Summary
Chapter 3 Mitosis detection in breast histology with fully fused and multi-scale fully fusedconvolution neural networks
3.1 Introduction
3.2 Pre-processing of the Hematoxylin & Eosin stained microscopy images
3.2.1 Stain-normalization of Hematoxylin & Eosin stained images
3.2.2 Patch extraction （coarsely extracted patches） from high resolution images
3.2.3 Patch extraction （fine discriminative patches） using sample selection strategy
3.3 Deep learning based methods
3.3.1 Features fused convolution neural network for mitosis detection
3.3.2 Multi-scale feature fused convolution neural network for mitosis detection
3.4 Experiments and results
3.4.1 Candidate detection on MITOS-ATYPIA-14 validation dataset using varioustypes of extracted patches
3.4.2 Visual results evaluated with FF-CNN and MFF-CNN models on validation HPFimages
3.4.3 Detection performance of FF-CNN and MFF-CNN models on test dataset
3.5 Summary
Chapter 4 Small mitotic cells detection using multi-scale object detector and atrous fullyconvolution based deep segmentation model
4.1 Introduction
4.2 Contributions
4.3 Publically available Mitosis datasets
4.4 Deep learning based Mitosis detection frameworks
4.4.1 Multi-scale region proposal model for mitotic cells detection
4.4.2 Atrous fully convolution model for bounding box estimation
4.5 Experimental analysis and results
4.5.1 Performance on ICPR 2012 grand challenge mitosis dataset
4.5.2 Performance on ICPR 2014 and AMIDA13 grand challenge mitosis dataset
4.6 Summary
Chapter 5 Multi-class breast cancer recognition with wavelet decomposed image basedconvolution neural network
5.1 Introduction
5.2 Contributions
5.3 Pre-processing of H&E stained histopathological images
5.3.1 Multi-class breast cancer histology datasets
5.3.2 Color normalization of breast histology images
5.3.3 Data augmentation with rotation technique
5.3.4 Data augmentation with proposed channel color augmentations algorithm
5.3.5 Theory of Wavelet transform
5.3.6 Haar wavelet decomposition of the breast histology images
5.4 Proposed deep convolution neural network based model for multi-class classification
5.5 Image classification techniques
5.5.1 Classification with Softmax classifier
5.5.2 Classification with support vector machine , k-nearest neighbor and Randomforest classifiers
5.6 Computational complexity of CNN models
5.6.1 Computations reduction of proposed HWDCNN models with Haar waveletdecomposed images
5.7 Experimental analysis and results
5.7.1 Performance on ICIAR 2018 multi-class dataset
5.7.2 Performance on Break His multi class cancer dataset
5.8 Summary
Conclusion
References
List of publications
Acknowledgements
Resume Details
本文编号：3007277
【文章来源】：哈尔滨工业大学黑龙江省 211工程院校 985工程院校
【文章页数】：156 页
【学位级别】：博士
【文章目录】：
摘要
Abstract
Acronyms and symbols
Chapter 1 Introduction
1.1 Background and significances of the study
1.2 Basic theory of deep learning
1.2.1 Deep learning
1.2.2 Deep learning programming frameworks
1.3 Steps involved in biopsy slide preparation and breast cancer grading
1.3.1 Fast slide scanners for digital image acquisition
1.3.2 Standard grading system for breast cancer recognition
1.4 Research status of Histopathology image detection techniques
1.4.1 Breast cancer detection/localization methods
1.4.2 Breast cancer segmentation methods
1.4.3 Breast cancer classification methods
1.5 Motivation
1.6 Limitations and challenges in breast histology image analysis
1.7 Thesis structure
1.8 Contributions
Chapter 2 Stain color normalization of hematoxylin and eosin stained histopathologyimages
2.1 Introduction
2.2 Stain color normalization of hematoxylin and eosin stained images
2.2.1 Benchmark color normalization methods to reduce color variation in breasthistology images
2.2.2 Proposed color normalization method based on image structure and color statics
2.3 Experimental results
2.3.1 Materials
2.3.2 Metrics to check the performance of color normalization methods
2.3.3 Experimental results and performance analysis
2.4 Summary
Chapter 3 Mitosis detection in breast histology with fully fused and multi-scale fully fusedconvolution neural networks
3.1 Introduction
3.2 Pre-processing of the Hematoxylin & Eosin stained microscopy images
3.2.1 Stain-normalization of Hematoxylin & Eosin stained images
3.2.2 Patch extraction （coarsely extracted patches） from high resolution images
3.2.3 Patch extraction （fine discriminative patches） using sample selection strategy
3.3 Deep learning based methods
3.3.1 Features fused convolution neural network for mitosis detection
3.3.2 Multi-scale feature fused convolution neural network for mitosis detection
3.4 Experiments and results
3.4.1 Candidate detection on MITOS-ATYPIA-14 validation dataset using varioustypes of extracted patches
3.4.2 Visual results evaluated with FF-CNN and MFF-CNN models on validation HPFimages
3.4.3 Detection performance of FF-CNN and MFF-CNN models on test dataset
3.5 Summary
Chapter 4 Small mitotic cells detection using multi-scale object detector and atrous fullyconvolution based deep segmentation model
4.1 Introduction
4.2 Contributions
4.3 Publically available Mitosis datasets
4.4 Deep learning based Mitosis detection frameworks
4.4.1 Multi-scale region proposal model for mitotic cells detection
4.4.2 Atrous fully convolution model for bounding box estimation
4.5 Experimental analysis and results
4.5.1 Performance on ICPR 2012 grand challenge mitosis dataset
4.5.2 Performance on ICPR 2014 and AMIDA13 grand challenge mitosis dataset
4.6 Summary
Chapter 5 Multi-class breast cancer recognition with wavelet decomposed image basedconvolution neural network
5.1 Introduction
5.2 Contributions
5.3 Pre-processing of H&E stained histopathological images
5.3.1 Multi-class breast cancer histology datasets
5.3.2 Color normalization of breast histology images
5.3.3 Data augmentation with rotation technique
5.3.4 Data augmentation with proposed channel color augmentations algorithm
5.3.5 Theory of Wavelet transform
5.3.6 Haar wavelet decomposition of the breast histology images
5.4 Proposed deep convolution neural network based model for multi-class classification
5.5 Image classification techniques
5.5.1 Classification with Softmax classifier
5.5.2 Classification with support vector machine , k-nearest neighbor and Randomforest classifiers
5.6 Computational complexity of CNN models
5.6.1 Computations reduction of proposed HWDCNN models with Haar waveletdecomposed images
5.7 Experimental analysis and results
5.7.1 Performance on ICIAR 2018 multi-class dataset
5.7.2 Performance on Break His multi class cancer dataset
5.8 Summary
Conclusion
References
List of publications
Acknowledgements
Resume Details
本文编号：3007277
本文链接：https://www.wllwen.com/kejilunwen/shengwushengchang/3007277.html
最近更新
教材专著
