Aggregated Wasserstein distance for hidden Markov models and automated morphological characterization of placenta from photos
Open Access
- Author:
- Chen, Yukun
- Graduate Program:
- Information Sciences and Technology
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- October 23, 2019
- Committee Members:
- James Z Wang, Dissertation Advisor/Co-Advisor
Jia Li, Committee Chair/Co-Chair
Xinyu Xing, Committee Member
Xiang Zhang, Committee Member
David Russell Hunter, Outside Member
Dashun Wang, Special Member
James Z Wang, Committee Chair/Co-Chair
Mary Beth Rosson, Program Head/Chair - Keywords:
- Data science
Aggregated Wasserstein
Optimal transport
Hidden Markov model
Placenta
Pathology
Morphological characterization
Deep learning - Abstract:
- In the past decade, fueled by the rapid advances of big data technology and machine learning algorithms, data science has become a new paradigm of science and has more and more emerged into its own field. At the intersection of computational methods, data modeling and domain knowledge, data science aims to solve a wide spectrum of problems with data-driven approaches. To really advance the field, a number of unique challenges are ahead of us. For example, from an algorithmic perspective, new fundamental modeling tools are needed to support innovative computing paradigms. And from a solving real-world problem perspective, novel data-driven solutions synergizing carefully-tailored cutting-edge models are yet to be developed for addressing meaningful problems not technology ready before, which include but not are limited to those in the fields of medical image analysis and diagnosis, climate change and weather forecasting. Within that big picture, the first part of this thesis proposed a new fundamental tool -- aggregated Wasserstein distances for hidden Markov models (HMMs) with state conditional distributions being Gaussian. By exploiting the fact that the marginal distribution for the type of HMMs in consideration at any time position follows a Gaussian mixture distribution (GMM), our proposed distances first softly register the states in two HMMs by solving an optimal transport optimization problem where the cost between components is the Wasserstein metric for Gaussian distributions. The solution of such optimization is a fast approximation to the Wasserstein metric between two GMMs. After state registration, our framework quantifies the dissimilarity of HMMs by measuring both the difference between the two marginal GMMs and that between the two transition matrices. Distances defined using our framework offer three key advantages. First, the defined distances are invariant to relabeling or permutation of states. Second, their definition is meaningful even for two HMMs that are estimated from data of different dimensionality, a situation that can arise due to missing variables. Third, they outperform Kullback-Leibler divergence (KLD)-based distances in terms of accuracy as well as efficiency in a variety of tasks, which is demonstrated through extensive experiments consisting of retrieval, classification and t-Distributed Stochastic Neighbor Embedding (t-SNE) visualization of time series, on both synthetic and real data. The second part of the thesis addressed an important medical image analysis problem: automated placental assessment and examination from photos. Specifically, we focus on morphological characterization, which includes the tasks of placental image segmentation, umbilical cord insertion point localization, and maternal/fetal side classification. To that end, we first curated a dataset consisting of approximately 1,300 placenta images taken at Northwestern Memorial Hospital, with hand-labeled pixel-level segmentation map, cord insertion point and other information extracted from the associated pathology report. Then we developed three encoder-decoder convolutional neural networks with a shared encoder to address those morphological characterization tasks by employing a transfer-learning training strategy. Through extensive experiments, we demonstrate that our method is able to produce accurate morphological characterization. We also show how our proposed method fits into a comprehensive two-stage placental assessment pipeline, which includes more placental feature analysis tasks, such as retained placenta (i.e., incomplete placenta), umbilical cord knot, meconium, abruption, chorioamnionitis, and hypercoiled cord, and categorization of umbilical cord insertion type. Our method and results for automated placental assessment may possess clinical impact and contribute to future pregnancy research. Moreover, this part of research is the first for comprehensive, automated, computer-based placental analysis and could serve as a launchpad for potentially multiple future innovations.