CASIA OpenIR  > 毕业生  > 硕士学位论文
COVID-19 CT影像的多尺度自适应分割算法研究








Other Abstract

At the end of 2019, the novel coronavirus COVID-19 began to emerge and quickly attracted the attention of the society. The outbreak of the epidemic has brought huge challenges to health systems around the world. COVID-19 is also evolving in the process of transmission, and new variants are also emerging. In particular, when COVID-19 is seriously infected, it will invade the lungs and cause pneumonia. As an effective method for screening pneumonia, computed tomography (CT) can provide interpretable visual information for the diagnosis and quantification of pneumonia.

In recent years, the deep learning algorithm has been more and more applied to the medical field, and has shown a strong auxiliary diagnosis ability in the diagnosis of many diseases. Therefore, the relevant methods based on deep learning can be used as an auxiliary means for diagnosis of pneumonia, can provide relevant visual intelligent diagnosis results, and can also play an effective role in monitoring the changes in the course of disease.

Based on the deep learning algorithm, this paper studies the adaptive segmentation and quantification method of the CT image of pneumonia caused by COVID-19 infection. The main research contents are as follows:

Relevant CT data were collected. The existing publicly labeled CT image data of pneumonia caused by COVID-19 infection were searched and sorted out, the relevant data distribution was analyzed, and a unified calculation format was established to provide data support for the subsequent in-depth learning algorithm.

A multi-scale segmentation model suitable for CT images is designed. Exploring a multi-scale segmentation baseline network based on encoder-decoder architecture, a hybrid balanced attention mechanism is designed to enhance the model's cognitive learning ability for lesion areas. An upsampling network is designed to enhance the decoder's learning ability. The overall effectiveness of the designed model was verified through comparison with existing traditional models and analysis of relevant modules. Meanwhile, the improved comparison basis is provided for the subsequent segmentation result analysis under different task settings.
A domain adaptation framework for CT images is designed to complete the domain adaptation of CT images from different sources. On the basis of the unsupervised Domain Adaptation adaptive migration framework based on image translation, the lesion-aware consistency constraint of lesion regions is introduced, forcing the deep learning model to focus on domain-invariant features, thus reducing the difference in domain distribution between private and public domains from the model perspective, greatly improving the segmentation performance of unlabeled private data, and greatly reducing the labeling dependency on private data. The visualization results prove the effectiveness of transfer learning.
Starting from CT image data, this paper explores how to improve the diagnostic ability of the algorithm based on deep learning for pneumonia caused by COVID-19 infection in terms of data processing, image segmentation and migration learning, and also provides new ideas for other applications that can use segmentation technology for artificial intelligence assisted diagnosis.

Keyword新型冠状病毒 CT 深度学习 图像分割 计算机辅助诊断
Indexed By其他
Sub direction classification医学影像处理与分析
planning direction of the national heavy laboratory多尺度信息处理
Paper associated data
Document Type学位论文
Recommended Citation
GB/T 7714
张振. COVID-19 CT影像的多尺度自适应分割算法研究[D],2023.
Files in This Item:
File Name/Size DocType Version Access License
Thesis.pdf(9077KB)学位论文 限制开放CC BY-NC-SA
Related Services
Recommend this item
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[张振]'s Articles
Baidu academic
Similar articles in Baidu academic
[张振]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[张振]'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.