Date of Degree

9-2020

Document Type

Dissertation

Degree Name

Ph.D.

Program

Computer Science

Advisor

Michael Grossberg

Committee Members

Zhigang Zhu

George Wolberg

Gig Mageras

Subject Categories

Artificial Intelligence and Robotics | Computer Sciences

Keywords

medical image segmentation, semi-automatic, conditional random field, graph cut, deep learning

Abstract

Medical image segmentation plays a crucial role in delivering effective patient care in various diagnostic and treatment modalities. Manual delineation of target volumes and all critical structures is a very tedious and highly time-consuming process and introduce uncertainties of treatment outcomes of patients. Fully automatic methods holds great promise for reducing cost and time, while at the same time improving accuracy and eliminating expert variability, yet there are still great challenges. Legally and ethically, human oversight must be integrated with ”smart tools” favoring a semi-automatic technique which can leverage the best aspects of both human and computer.

In this work we show that we can formulate a semi-automatic framework for the segmentation problem by formulating it as an energy minimization problem in Conditional Random Field (CRF). We show that human input can be used as adaptive training data to condition a probabilistic boundary term modeled for the heterogeneous boundary characteristics of anatomical structures. We demonstrated that our method can effortlessly adapt to multiple structures and image modalities using a single CRF framework and tools to learn probabilistic terms interactively. To tackle a more difficult multi-class segmentation problem, we developed a new ensemble one-vs-rest graph cut algorithm. Each graph in the ensemble performs a simple and efficient bi-class (a target class vs the rest of the classes) segmentation. The final segmentation is obtained by majority vote. Our algorithm is both faster and more accurate when compared with the prior multi-class method which iteratively swaps classes. In this Thesis, we also include novel volumetric segmentation algorithms which employ deep learning and indicate how to synthesize our CRF framework with convolutional neural networks (CNN). This would allow incorporating user guidance into CNN based deep learning for this task. We think a deep learning based method interactively guided by human expert is the ideal solution for medical image segmentation.

Share

COinS