基于神经网络的肺栓塞和肺结节的分割和分类算法研究

发布时间：2021-02-23 03:21

　　肺栓塞（Pulmonary Embolism,PE）是人类与癌症相关的死亡的最常见原因之一。用于医学疾病筛查的计算机断层扫描（Computer Tomography,CT）是肺栓塞敏感性高且早期发现的无创诊断方法,可大大提高生存率。但是,解释医学图像并制定评估或护理决策需要专门合格的医学专家,当前解释诊断图像的方法是费力、费时、昂贵且容易出错的。因此,基于神经网络模型的辅助诊断具有重要意义,该模型将自动提供诊断建议。深度学习的最新发展鼓励我们重新考虑临床诊断专注于医学图像的方式。事实证明,早期发现对于为患者提供最大的康复和生存可能性至关重要。在本文中,我们提出了一种基于CT图像的神经网络框架,对肺栓塞和结节进行了全自动分割和分类。我们的工作包括两部分:PE分割（预处理和训练模型以进行PE分割）和分类（将候选结节诊断和分类为良性或恶性）。对于PE分割,我们结合CT窗口技术和图像裁剪,设计了一种新的有效图像预处理方法。建立的该模型是编码器-解码器卷积网络,剩余块代替原始卷积块用于U-Net。为了进行分类,奇异结节,分别构建了两个深层3D Conv Nets用于结节检测和分类。此外,我们验证... 

【文章来源】：西南科技大学四川省

【文章页数】：63 页

【学位级别】：硕士

【文章目录】：
摘要（Chinese Abstract）
ABSTRACT 英文摘要
1.Introduction
    1.1 Overview
    1.2 Dissertation Outline and Contribution
2.Literature Review
    2.1 Background
    2.2 Overview
    2.3 Popular Algorithms
        2.3.1 Supervised Machine Learning Algorithms
        2.3.2 Unsupervised Machine Learning Algorithms
        2.3.3 Semi-supervised Machine Learning Algorithms
        2.3.4 Reinforcement Machine Learning Algorithms
        2.3.5 Recommender Systems
        2.3.6 Deep Learning
3.Segmentation of Pulmonary Embolism
    3.1 Segmentation of Pulmonary Embolism using Neural Networks
    3.2 Related Works
    3.3 Methodology
        3.3.1 Dataset
        3.3.2 Data Preparation
        3.3.3 Image Cropping
        3.3.4 Window Technique
        3.3.5 Standard Normalization
        3.3.6 Image augmentation
        3.3.7 Image Post-Processing
        3.3.8 Evaluation
    3.4 Experiments
        3.4.1 System Specification and tools
    3.5 Results and Discussion
4.Classification of Pulmonary Nodules
    4.1 Classification of Pulmonary Nodules using Neural Networks
    4.2 Related Work
    4.3 Methodology
        4.3.1 Neural Network Framework for Nodule Detection
        4.3.2 3D Faster R-CNN with Deep3D Dual Path Net for Nodule Detection
        4.3.3 Gradient Boosting Machine3D Dual Path Net Function for Nodule Classification
        4.3.4 Neural Network for Fully Automated PE CT Nodules Diagnosis
    4.4 Experiments
        4.4.1 Datasets
        4.4.2 System Specification and tools
        4.4.3 Preprocessing
    4.5 Results
        4.5.1 Neural network for Nodule Detection
        4.5.2 Neural network for Nodule Classification
        4.5.3 Compared to experienced physicians on their individual positive nodules
    4.6 Discussion
        4.6.1 Nodule Detection
        4.6.2 Classification of Nodules
5.Conclusion and Future work
    5.1 Conclusion
    5.2 Future Work
Acknowledgements
References
Achievements
Research achievements during the Undergraduate degree



本文编号：3046935