Dissertation
Deep learning-based pulmonary image registration and analysis
University of Iowa
Doctor of Philosophy (PhD), University of Iowa
Spring 2023
DOI: 10.25820/etd.007233
Abstract
Pulmonary image registration is important because it can be used to compute lung biomechanical measures, study disease progression, and guide image surgery or radiotherapy. The learning-based pulmonary image registration method for aligning 3D breath-hold inspiration-expiration images is a challenging problem due to its large deformation magnitude, the change in CT intensity across the breathing cycle, and anatomical variations among subjects. The overall objective of this thesis is to develop and implement a deep learning-based registration algorithm to align 3D pulmonary CT images with large deformation magnitudes and investigate the regional lung mechanics and longitudinal progression pattern of chronic obstructive pulmonary disease (COPD) with image registration. We designed a fast unsupervised learning-based framework, named Population Learning followed by One Shot Learning (PLOSL), for the 3D-CT pulmonary image registration algorithm. This registration algorithm combines population learning (PL) and one-shot learning (OSL) together, which has the advantages of the PL and OSL approaches while reducing their respective drawbacks. A coarse-to-fine convolution encoder-decoder CNN architecture is used to register large and small shape features. PLOSL pulmonary image registration uses tissue volume preserving and vesselness constraints to register inspiration-to-expiration and expiration-to-inspiration pulmonary CT images.
The second objective of this research was to validate and evaluate the registration performance of the proposed method PLOSL in registering intra-patient chest CT images. Results demonstrated PLOSL could achieve comparable registration performance to conventional iterative image registration algorithms in terms of landmark error while preserving pulmonary topology on multiple datasets. Results also show that the proposed PLOSL can generate robust and consistent lung biomechanical measures as conventional iterative image registration algorithms across varying disease severity levels while increasing efficiency by 1.5 orders of magnitude.
The third aim of this research was to investigate the longitudinal progression pattern of CT-quantified outcomes of COPD at the voxel level. The correlation of regional lung function decline with emphysema and air-trapping progression was explored. Results show that regional lung function decline was correlated with increased air-trapping progression. The voxel-wise progression pattern of emphysema and functional small airway disease (fSAD) was analyzed across time as a function of the GOLD stage. Results show fSAD progression precedes the development of emphysema and the lung region affected by fSAD was at high risk of developing emphysema. Furthermore, we investigated the utility of the deep learning model for predicting the local progression of COPD for continuous progressors. Results show that CNN was able to predict the future local-level COPD progression (i.e., emphysema and air-trapping) for continuous progressors based on the earlier CTs.
Details
- Title: Subtitle
- Deep learning-based pulmonary image registration and analysis
- Creators
- Di Wang
- Contributors
- Gary Edward Christensen (Advisor)Joseph M Reinhardt (Committee Member)Oguz Durumeric (Committee Member)Eric A Hoffman (Committee Member)Punam K Saha (Committee Member)
- Resource Type
- Dissertation
- Degree Awarded
- Doctor of Philosophy (PhD), University of Iowa
- Degree in
- Electrical and Computer Engineering
- Date degree season
- Spring 2023
- Publisher
- University of Iowa
- DOI
- 10.25820/etd.007233
- Number of pages
- xvii, 195 pages
- Copyright
- Copyright 2023 Di Wang
- Language
- English
- Date submitted
- 04/20/2023
- Date approved
- 06/30/2023
- Description illustrations
- color illustrations
- Description bibliographic
- Includes bibliographical references (pages 184-195).
- Public Abstract (ETD)
Pulmonary image registration is important because it can be used to study the biomechanical properties of lung and lung disease. The first contribution of this dissertation is a deep learning-based registration algorithm, named Population Learning followed by One Shot Learning (PLOSL), which was developed to align 3D pulmonary CT images using less computational resources than traditional iterative image registration algorithms.
The second contribution is a study comparing the PLOSL algorithm to four conventional iterative image registration algorithms. Results demonstrated that PLOSL could achieve comparable registration performance to conventional image registration algorithms in terms of landmark error while preserving pulmonary topology on multiple datasets. Results also show PLOSL can generate robust and consistent lung biomechanical measures as conventional algorithms while reducing the computation time by 1.5 orders of magnitude.
The third contribution of this dissertation is a study of the voxel-wise longitudinal progression pattern of CT-quantified outcomes of chronic obstructive pulmonary disease (COPD) using PLOSL image registration. Data from the SPIROMICS study were analyzed. Results show that functional small airway disease (fSAD) progression precedes the development of emphysema and the lung region affected by fSAD was at high risk of developing emphysema. The fourth contribution is a deep learning model that was proposed to predict the local-level COPD progression for continuous progressors based on the earlier CTs. Results show that the proposed model is effective for predicting the future extent and spatial distribution of COPD progression.
- Academic Unit
- Electrical and Computer Engineering
- Record Identifier
- 9984437258802771
Metrics
3 File views/ downloads
22 Record Views