Dissertation
Deep learning meets graph: novel hybrid methods for improved quantitative medical image analysis
University of Iowa
Doctor of Philosophy (PhD), University of Iowa
Autumn 2019
DOI: 10.17077/etd.005235
Abstract
Medical image analysis plays a vital role in medical diagnostics, helping physicians monitor disease progression, select the most proper treatment plans in clinical practice, and in many other cases. In various image analysis tasks, image segmentation often serves as the first step. Later on, measurements extracted from segmented regions like intensity distribution, volume, and boundary properties provide a straightforward approach to inspect and interpret pathology. When rendered in 2D/3D/4D, the segmentations reveal non-trivial topological information. Furthermore, the extracted and quantized features can be combined with medical records to perform functional predictions especially when trained on data from large populations. However, traditional manual segmentation often requires topical expertise, suffers from inter- and intra-observer variations, and is overly time-consuming. Many semi-automated and automated algorithms have been proposed to reduce humans' interaction when generating accurate image segmentations, including active shape models, atlas based methods, graph based methods, and artificial neural networks. Among these methods, graph based methods have solid mathematical proof of global optimality and can be easily generalized to work on n-dimensional images, while they lack a good feature extractor to guide the necessary graph construction and define the underlying cost metrics. Recently, deep learning based methods have become prevalent and achieved new levels of the state-of-the-art accuracy in medical image segmentation. However, they are not explicitly topology-wise and require a large amount of annotations during training.
Details
- Title: Subtitle
- Deep learning meets graph: novel hybrid methods for improved quantitative medical image analysis
- Creators
- Zhihui Guo
- Contributors
- Milan Sonka (Advisor)Joseph M Reinhardt (Committee Member)Sajan Goud Lingala (Committee Member)Xiaodong Wu (Committee Member)Reinhard R Beichel (Committee Member)Andreas Wahle (Committee Member)
- Resource Type
- Dissertation
- Degree Awarded
- Doctor of Philosophy (PhD), University of Iowa
- Degree in
- Biomedical Engineering
- Date degree season
- Autumn 2019
- DOI
- 10.17077/etd.005235
- Publisher
- University of Iowa
- Number of pages
- xiv, 114 pages
- Copyright
- Copyright 2019 Zhihui Guo
- 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
- color illustrations
- Description bibliographic
- Includes bibliographical references (pages 104-114).
- Public Abstract (ETD)
Medical image analysis plays a vital role in helping physicians arrive at a diagnosis, monitor disease progression, and select proper treatment plans in clinical practice. In various image analysis tasks, image segmentation often serves as the first step. From the segmented regions, information like region size, location, and shape properties can be extracted for quantitative analysis. Traditionally, clinicians need to annotate where the pathologies are or trace tissue contours on images slice by slice manually. This kind of manual segmentation is tedious and time-consuming. In contrast, computer-aided image segmentation approaches can reduce humans’ efforts and generate image segmentations automatically.
In this work, we study novel combinations of two sets of automated image segmentation approaches, graph optimization and deep learning, to improve accuracy in medical image analysis. A graph is a mathematical model with nodes and pairwise connections that can translate an image segmentation task into a cost minimizing problem. Though global and local shape constraints are precisely encoded, the construction of a graph usually depends on human-selected features. A deep learning classifier learns desired patterns from raw image input without hand-crafted features, while a large training set is needed to obtain high accuracy and the prediction is at the pixel-level. By introducing deep learning into graph construction, we strengthen the advantages of both methods and compensate for their weaknesses. To demonstrate effectiveness of such combinations, experiments were done for segmentation tasks from normal tissue to lesions on computed tomography (CT) and magnetic resonance (MR) images. Besides image segmentation, we also study medical outcome prediction with the help of both graph based and deep learning based methods. Prediction of Humphrey visual function from structural information obtained from segmented retinal images is presented for glaucoma patients as one such example.
- Academic Unit
- Roy J. Carver Department of Biomedical Engineering
- Record Identifier
- 9983779800002771
Metrics
86 File views/ downloads
119 Record Views