Dissertation
Deep convolutional neural network based analysis methods for radiation therapy applications
University of Iowa
Doctor of Philosophy (PhD), University of Iowa
Autumn 2022
DOI: 10.25820/etd.006662
Abstract
Medical imaging is crucial for many modern clinical applications such as diagnostics, disease progression monitoring, and treatment planning. For a large number of these applications, image segmentation serve as the first processing step, followed by quantitative image feature extraction to assess size, volume, shape characteristics, and intensity-related features. These quantitative features provide valuable information for interpretation of pathology and prognostics. However, the segmentation process is mostly done manually, which is time-consuming and lacks consistency. Automated and semi-automated methods have been proposed over the years to increase the efficiency, consistency, and accuracy of segmentation. Among the many methods, there are simple ones like thresholding, region-growing, watershed and more advanced ones like active contour methods, active shape model, and more recently, deep learning-based methods. Deep learning models have yielded a new generation of image segmentation methods with remarkable performance improvements, typically beating standard and traditional machine learning methods in various benchmarks. Another branch of medical image analysis is radiomics. A radiomics approach refers to methods that extracts features from medical images using data-characterisation algorithms, which can be utilized to discover and quantify characteristics of tumors. In this field, deep learning-based approaches are also appearing, showing promising improvements compared to using handcrafted features.
This thesis aims to investigate and develop deep learning-based methods for the use in radiation oncology to facilitate the efficient analysis of volumetric medical image datasets. In particular, the segmentation performance of several state-of-the-art convolutional neural networks (CNN) and more recent transformer-based models for segmenting both normal and abnormal structures in PET images were studied and compared. For cerebellum segmentation, we discovered a modified U-Net that has significantly better segmentation performance than the original U-Net and V-Net. For head and neck cancer (HNC) lesion segmentation, we proposed a U-Net-CBAM that slightly outperforms the U-Net and several state-of-the-art transformer-based models. We also proposed a combined localization and segmentation approach for efficient segmentation of the pelvis in CT images with variant volume sizes and image qualities, achieving up to four times faster processing speed compared to a standard approach. Furthermore, this single object combined localization and segmentation approach is extended to process multiple objects in a scan at once. Specifically, the algorithm is applied and studied in the context of lumbar and thoracic vertebrae segmentation. In addition, we explored the performance of using CNN as well as a transformer-based models for HNC outcome prediction and found that the transformer-based model shows the best prognostic performance.
In conclusion, our joint localization and segmentation approach for pelvis and lumbar and thoracic vertebrae demonstrated a novel way of utilizing CNNs to substantially improve the segmentation efficiency. We assessed the segmentation performance difference of several CNN and transformer based models on two different data sets. The difference can be substantial. Therefore, testing different models before starting a new project is imperative. In addition, comparison between CNN models and transformer based models on HNC outcome prediction performance showed the potential of transformer based models for radiomics applications. Furthermore, annotation quality, data size and class balance were found to play an important role in deep learning-based segmentation and classification performance, respectively.
Details
- Title: Subtitle
- Deep convolutional neural network based analysis methods for radiation therapy applications
- Creators
- Xiaofan Xiong
- Contributors
- Reinhard R Beichel (Advisor)Joseph M Reinhardt (Committee Member)Terry A Braun (Committee Member)Brian J Smith (Committee Member)John J Sunderland (Committee Member)John M Buatti (Committee Member)
- Resource Type
- Dissertation
- Degree Awarded
- Doctor of Philosophy (PhD), University of Iowa
- Degree in
- Biomedical Engineering
- Date degree season
- Autumn 2022
- Publisher
- University of Iowa
- DOI
- 10.25820/etd.006662
- Number of pages
- xvii, 140 pages
- Copyright
- Copyright 2022 Xiaofan Xiong
- Comment
This thesis has been optimized for improved web viewing. If you require the original version, contact the University Archives at the University of Iowa: https://www.lib.uiowa.edu/sc/contact/.
- Language
- English
- Description illustrations
- illustrations, tables, graphs
- Description bibliographic
- Includes bibliographical references (pages 127-140).
- Public Abstract (ETD)
Medical image analysis is important in clinical applications that helps clinicians with diagnostics, disease progression monitoring, and treatment planning. For a large number of these tasks, image segmentation serves as the first processing step and enables image feature extraction to assess region size, texture, and shape properties. For example, these features provide valuable information for quantitative analysis of pathology. The segmentation process is traditionally done by hand, requiring the physicians to manually draw the target structures using a computer. This is a tedious process and the results are prone to errors and inter- as well as intra-user variability. To solve this problem, automated segmentation methods are needed that are more efficient and can generate required segmentations more accurately and consistently. This facilitates the extraction of quantitative image features. In addition, the features can be utilized to discover and quantify characteristics of tumors, which can be used to predict the outcome of cancer patients.
In this thesis, we investigated the performance of several deep learning-based image segmentation algorithms in segmenting different normal and abnormal structures in different types of medical images. We discovered an algorithm that can better segment cerebellum in PET images compared to other more popular algorithms. We proposed a new algorithm that outperforms some advanced current algorithms in segmenting lesions for head and neck cancer patients. We also proposed a novel localization and segmentation approach that makes segmenting pelvic bone in computed tomography (CT) images more efficient, achieving up to four times faster processing speed compared to a standard approach. This approach is further extended to be applicable to a part of the spine that includes up to 18 vertebrae. In addition, we explore the performance of using several algorithms for predicting the treatment outcome of patients with head and neck cancer (HNC) and discovered that one of the algorithms showed better prediction results.
In conclusion, our joint localization and segmentation approach for pelvis as well as lumbar and thoracic vertebrae extended the possibility of significantly improving the segmentation efficiency of deep learning-based methods. We found that testing different deep learning models before starting a new project can be beneficial. In addition, we demonstrated the potential of a new deep learning algorithm for predicting HNC outcome.
- Academic Unit
- Roy J. Carver Department of Biomedical Engineering
- Record Identifier
- 9984362858602771
Metrics
4 File views/ downloads
40 Record Views