Dissertation
Globally optimal surface segmentation for medical images using deep learning: algorithms, robustness and applications
University of Iowa
Doctor of Philosophy (PhD), University of Iowa
Summer 2020
DOI: 10.17077/etd.005508
Abstract
Automated image segmentation plays a very import role in quantitative image analysis. In recent several years, deep learning (DL) or CNN based method for semantic segmentation has become very popular in computer vision and then in medical image research communities.
Most deep learning based semantic segmentation methods are region-based, i.e. each pixel is labeled as target object or background. As there is no explicit constraint among the labeling of different pixels, usually post processing, such as morphological operation, is required to get reasonable prediction.
Actually, one can also model the segmentation problem as surface based: each pixel is labeled as surface or non-surface and then target object can be defined based on prior knowledge, i.e. one has to know region on which side of surface corresponds to target object. Apparently these two views are equivalent but modeled in different way. As one prominent surface-based method, Graph-Search (GS) has achieved great success, especially in medical field. This method is capable of simultaneously detecting multiple interacting surfaces, in which the optimality is controlled by the cost functions designed for individual surfaces and by several geometric constraints defining the surface smoothness and interrelations. However, the cost function design is usually non-trivial.
Inspired by the GS, one research focus is to solve the surface segmentation problem with global optimality guarantee for both 3-D and 2-D images using deep learning, so that the cost functions are learned by deep networks utilizing training dataset.
In the era of big data, automated deep learning based image processing enables to process a large amount of medical image data efficiently. This is especially helpful for semantic segmentation, as manual contouring is tedious and time consuming. In many applications, the deep learning based segmentation methods can even achieve expert-level accuracy. However, in practice, deep learning methods may fail due to many factors: such as domain shift, adversarial noise, and low image quality. Therefore, another research focus is testing the robustness of deep learning based segmentation methods and even predicting segmentation quality without ground truth.
To accomplish the first objective of this thesis work, novel graph based frameworks are proposed to seamlessly integrated with deep neural networks. First, a graph model based on convex quadratic programming is developed, the solving of which is guaranteed globally optimal. The developed model is applied and validated on the intra-retinal layer segmentation of Spectral Domain Optical Coherence Tomography (SD-OCT) images of eye and vessel walls in Intravascular Ultrasound (IVUS) images. Second, to extend this model based deep learning surface segmentation method to 3-D, one shape-aware patch generation approach for Genus-0 surface is proposed to enable normal 3-D CNN to properly work on. The developed method is applied to prostate segmentation of MR images and spleen segmentation of CT images in 3-D. Third, the robustness of the proposed model based surface segmentation method is tested under adversarial attacks and one robust segmentation quality assessment approach based on conditional reconstruction networks is also developed and verified on cardiac left-ventricular myocardium (LVM) segmentation.
Details
- Title: Subtitle
- Globally optimal surface segmentation for medical images using deep learning: algorithms, robustness and applications
- Creators
- Leixin Zhou
- Contributors
- Xiaodong Wu (Advisor)Mona K Garvin (Committee Member)Weiyu Xu (Committee Member)Mathews Jacob (Committee Member)Stephen Baek (Committee Member)
- Resource Type
- Dissertation
- Degree Awarded
- Doctor of Philosophy (PhD), University of Iowa
- Degree in
- Electrical and Computer Engineering
- Date degree season
- Summer 2020
- DOI
- 10.17077/etd.005508
- Publisher
- University of Iowa
- Number of pages
- xvi, 125 pages
- Copyright
- Copyright 2020 Leixin Zhou
- Language
- English
- Description illustrations
- color illustrations
- Description bibliographic
- Includes bibliographical references (pages 112-125).
- Public Abstract (ETD)
Automated image boundary/surface segmentation plays a very import role in quantitative image analysis. In recent several years, deep learning (DL) or CNN based method for surface segmentation has become very popular in computer vision and then in medical image research communities. In many applications, the deep learning based segmentation methods can even achieve expert-level accuracy. However, in practice, deep learning methods may fail due to many factors: such as domain shift [51], adversarial noise [2], and low image quality. In some sense, most deep learning based semantic segmentation methods are still black-box like, where usually lacks of guarantee. Prior to current deep learning era, to improve the performance, usually prior information, e.g. smoothness, is injected to traditional model based methods.
In this doctoral thesis, inspired by Graph-Search (GS) [77, 40] method, novel graph model based deep learning surface framework is proposed to improve the surface segmentation performance and especially robustness. Human designed priors and deep network learned priors both can be integrated to the model seamlessly. The proposed framework achieves better segmentation accuracy and robustness. The other focus of this thesis is to predict when the segmentation method fails, and then human experts can get involved. For this sake, one reconstruction based segmentation quality assessment method is proposed based on conditional generative networks. The proposed method is validated to be more robust than state-of-the-art method.
- Academic Unit
- Electrical and Computer Engineering
- Record Identifier
- 9983987895102771
Metrics
55 File views/ downloads
150 Record Views