Dissertation
Deep learning of lung diseases on computed tomography images
University of Iowa
Doctor of Philosophy (PhD), University of Iowa
Spring 2023
DOI: 10.25820/etd.007237
Abstract
This research is a collection of three papers on the use of deep learning to detect and classify lung diseases or lung injuries using a CAE-FC model or a contrastive learning model. Due to a broad spectrum of clinical presentations for a complex lung disease, its pathogenesis remains unclear or unsettled. As a result, a general descriptive term such as chronic obstructive pulmonary disease (COPD) is still used. Identification of new classifications that reflect the phenotypes of these diseases is of great interest. In my research work, two deep learning models were developed to extract computed tomography (CT) imaging features for cluster analysis and identify subtypes of different lung malfunctions as indicated in the following three research projects. Initially, an unsupervised three-dimensional (3D) convolutional autoencoder (CAE)-feature constructor (FC) deep learning network was built to extract imaging features of lung tissue from two-channel images, and then a novel contrastive learning model with a lung volume transform was utilized to learn latent features of disease phenotypes from CT scans at inspiration and expiration of the same subjects. The major findings and conclusions for each study are summarized below:With the first research project, two important latent traits, factor 0 (F0) and factor 4 (F4), among 541 former smokers with COPD were identified. F0 and F4 were sufficient to identify healthy non-smokers from subjects at risk and differentiate COPD patients at all levels of spirometry-staged severity. Because of their detectability of abnormal tissue patterns, F0 and F4 could potentially serve as surrogate markers for new COPD phenotypes either in detecting early abnormalities in susceptible subjects at risk or in assessing the rate of lung function decline in severe patients.
With the second research project, two subject-clusters, cluster 0 (C0) and cluster 5 (C5) were identified from 96 subjects who had been exposed to toxic humidifier disinfectants (HD), although their lung CT and PFT appeared normal. Subjects in C5 were considered as subjects not affected by HD. On the other hand, subjects in C0 were characterized by increased lung tissue and hypo-inflation, in comparison with subjects in C5, suggesting that C0 subjects may be at risk of HD-induced inflammation or lung fibrosis. Moreover, with CFPD analysis, C0 subjects tend to have higher airway resistance and HD aerosol particle deposition density, which may be related to the susceptibility to HD-induced lung damage.
With the third research project, two clinically meaningful subtypes among 140 post-COVID-19 subjects were identified. Cluster 1 (C1) and cluster 2 (C2) are characterized by subjects with air-trapping caused by small airways disease and subjects with airway-associated interstitial fibrotic-like patterns, respectively.
The results shown above indicate that the proposed deep learning models can facilitate the recognition and interpretation of these aforementioned subtypes and can eventually provide an effective customized guidance to patient’s healthcare. Hopefully, an automatic detection system for medical care can be developed accordingly in the near future.
Details
- Title: Subtitle
- Deep learning of lung diseases on computed tomography images
- Creators
- Frank Li
- Contributors
- Ching-Long Lin (Advisor)Alejandro P Comellas Freymond (Committee Member)Eric A Hoffman (Committee Member)Joseph M Reinhardt (Committee Member)Suresh Raghavan (Committee Member)
- Resource Type
- Dissertation
- Degree Awarded
- Doctor of Philosophy (PhD), University of Iowa
- Degree in
- Biomedical Engineering
- Date degree season
- Spring 2023
- Publisher
- University of Iowa
- DOI
- 10.25820/etd.007237
- Number of pages
- xv, 144 pages
- Copyright
- Copyright 2022 Frank Li
- Language
- English
- Date submitted
- 01/23/2023
- Date approved
- 05/24/2023
- Description illustrations
- color illustrations
- Description bibliographic
- Includes bibliographical references (pages 132-144).
- Public Abstract (ETD)
It is crucial to identify the phenotypes of the complex lung disease for disease management, pathogenesis research, and drug development. This thesis utilized two deep learning models to extract computed tomography (CT) imaging features for cluster analysis. Our developed deep learning models were able to identify subtypes of three different lung malfunctions.
First, two important latent traits, factor 0 (F0) and factor 4 (F4), among former smokers with COPD, were identified and could be used for new COPD phenotype identification, either in detecting an early abnormality in susceptible subjects at risk or in assessing the rate of lung function decline in severe patients, respectively.
Second, two subject-clusters, cluster 0 (C0) and cluster 5 (C5), were identified from subjects who had been exposed to toxic humidifier disinfectants (HD), although their lung CT and PFT appeared normal. C5 was a factor characterized by subjects not affected by HD. On the other hand, subjects in C0 may be at risk of HD-induced lung damage.
Finally, two clinically meaningful subtypes among the post COVID-19 subjects were identified. Cluster 1 (C1) and cluster 2 (C2) are characterized by subjects with air-obstruction caused by small airways narrowing and subjects with airway-associated injuries, respectively.
The previous results indicate that the proposed deep learning models can serve as an effective tool for facilitating the recognition and interpretation of these aforementioned subtypes so that an effective guidance to patient’s healthcare can be customized. Hopefully, an automatic detection system for clinical uses can be developed accordingly in the near future.
- Academic Unit
- Roy J. Carver Department of Biomedical Engineering
- Record Identifier
- 9984428943202771
Metrics
16 File views/ downloads
7 Record Views