CASIA OpenIR  > 毕业生  > 硕士学位论文
基于深度学习的特发性肺纤维化HRCT影像特征识别与分割
朱志敏
Subtype硕士
Thesis Advisor郭大蕾
2020-06-03
Degree Grantor中国科学院大学
Place of Conferral中国科学院自动化研究所
Degree Discipline控制工程
Keyword深度学习,医学图像分类,特发性肺纤维化,医学图像分割,计算机辅助诊断
Abstract

近年来,深度学习算法在疾病诊断中的应用达到了前所未有的高度和规模,已成为医学图像分析和疾病辅助诊断的一种新方法。特发性肺纤维(idiopathic pulmonary fibrosis,IPF)是一种“类肿瘤疾病”,其重要诊断依据是肺部高分辨率计算机断层扫描(high-resolution computed tomography,HRCT)的典型影像学特征。基于深度学习的HRCT无创、可重复和局部性的计算机分析方法,是探索辅助诊断IPF疾病的基础和必要环节。

本文基于深度学习算法对特发性肺纤维化的影像特征进行分类识别与分割量化研究,主要内容如下:

研究了IPF最主要病灶影像特征——蜂窝影的分类问题,分别设计了基于经验模型、基于网络结构随机搜索模型、基于经典分类网络模型和基于投票法的集成模型来对IPF蜂窝影病灶进行分类识别。通过对比和分析各模型的性能得出了最佳分类模型,也为后续分割和小样本研究结构改进时提供了模型选择的参考依据。

研究了IPF典型病灶的分割问题,分别以U-Net和U-Net++为基础模型,对IPF典型病灶分割量化的最佳模型进行探索。实验结果显示超越了其他用于肺部病灶分割模型的性能,并且从中发现了病灶边缘信息对于分割结果的重要性。

研究了基于小样本学习的IPF全种类病灶分类问题,用关系网络模型解决肺部部分病灶数据量不足的问题,并尝试了对该模型的特征提取网络的改进研究。结果表明改进后的算法能有效地对IPF全种类病灶进行分类识别。最后,给出了ILDs数据集中的病例的IPF最终诊断结果示例。

本文从数据处理、迁移学习、模型结构改进、小样本学习等方面提升了基于深度学习的特发性肺纤维化HRCT影像特征识别与分割效果,为建立特发性肺纤维化计算机辅助诊断系统提供了新方法和新思路。

Other Abstract

In recent years, the application of deep learning algorithms in disease diagnosis has reached an unprecedented height and scale, and has become a new method for medical image analysis and disease auxiliary diagnosis. Idiopathic pulmonary fibrosis (IPF) is a kind of "tumor-like" disease, and its important diagnosis is based on the high-resolution computed tomography (HRCT) of the lung. Therefore, looking for a non-invasive, repeatable, and local analysis method for HRCT based on deep learning has been the basis and necessary link for computer-aided diagnosis of IPF disease.

In this paper, we study the recognition and segmentation of the image features of idiopathic pulmonary fibrosis by deep learning algorithms. The main contents are as follows:

Classification of the most important lesions in IPF. We build classification algorithms based on experience model, random search model based on network structure, classic classification network model and voting-based integration model to distinguish honeycombing from other lesions. By comparing and analyzing the performance of each model, we finally get the best classification model, which also provides a reference for the research in the next two chapters.

Segmentation of IPF typical lesions. We explore the optimal segmentation model of IPF typical lesions based on U-Net and U-Net++ structure. The results are charming as the performance of our model surpasses the previous results in segmentation. The results also draw our attention to the importance of edge information in segmentation problems.

Classification of all seventeen kinds of IPF lesions based on few-shot learning (FSL). We use the relational network (RN) model to solve the data shortage problem of certain lung lesions. On top of that, we try to improve the feature extraction network of the model. The results show that the improved algorithm can identify and classify all types of IPF lesions effectively. We give the final IPF diagnosis results of the cases in the ILDs dataset at the end of the paper.

In this paper, we improve the performance of HRCT image feature recognition and segmentation based on deep learning by using data processing, transfer learning, model structure improvement, few-shot learning, and so on, which provides a new way for the establishment of computer-aided diagnosis system of idiopathic pulmonary fibrosis.

Pages85
Language中文
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/39879
Collection毕业生_硕士学位论文
Recommended Citation
GB/T 7714
朱志敏. 基于深度学习的特发性肺纤维化HRCT影像特征识别与分割[D]. 中国科学院自动化研究所. 中国科学院大学,2020.
Files in This Item:
File Name/Size DocType Version Access License
最终版_硕士毕业设计_基于深度学习的特发(2686KB)学位论文 暂不开放CC BY-NC-SAApplication Full Text
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[朱志敏]'s Articles
Baidu academic
Similar articles in Baidu academic
[朱志敏]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[朱志敏]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.