Master of Engineering Science
Medical object detection and segmentation are crucial pre-processing steps in the clinical workflow for diagnosis and therapy planning. Although deep learning methods have achieved considerable performance in this field, they impose several shortcomings, such as computational limitations, sub-optimal parameter optimization, and weak generalization. Deep reinforcement learning as the newest artificial intelligence algorithm has great potential to address the limitation of traditional deep learning methods, as well as obtaining accurate detection and segmentation results. Deep reinforcement learning has a cognitive-like process to propose the area of desirable objects, thereby facilitating accurate object detection and segmentation. In this thesis, we deploy deep reinforcement learning into two challenging and representative medical object detection and segmentation tasks: 1) Sequential-Conditional Reinforcement Learning (SCRL) for vertebral body detection and segmentation by modeling the spine anatomy with deep reinforcement learning; 2) Weakly-Supervised Teacher-Student network (WSTS) for liver tumor segmentation from the non-enhanced image by transferring tumor knowledge from the enhanced image with deep reinforcement learning. The experiment indicates our methods are effective and outperform state-of-art deep learning methods. Therefore, this thesis improves object detection and segmentation accuracy and offers researchers a novel approach based on deep reinforcement learning in medical image analysis.
Summary for Lay Audience
Automatic medical object detection and segmentation based on artificial intelligence as computer-assisted-diagnosis tools are significant for clinicians in the disease diagnosis and treatment planning. Medical object detection and segmentation distinguish the object of interest from the medical image, which provides clinicians with the location, shape, and size of the object, thereby assisting clinicians to make a decision. Deep learning methods have achieved considerable performance in this field by leveraging convolutional neural networks. However, as the development of deep learning, it also imposes some limitations and its accuracy in some tasks cannot meet clinical expectations. In this case, this thesis seeks to employ deep reinforcement learning to address the limitations of deep learning methods and obtain accurate medical object detection and segmentation. Particularly, this thesis deploys deep reinforcement learning in vertebral body segmentation, where the newly-proposed Sequential-Conditional Reinforcement Learning (SCRL) models the spine anatomy as a sequential decision-making process and segments vertebral bodies along the spine. In another project, this thesis deploys deep reinforcement learning into a more challenging task. Particularly, this thesis proposes the Weakly-Supervised Teacher-Student network (WSTS) to address liver tumor segmentation from the non-contrast-enhanced image. WSTS leverages deep reinforcement learning to transfer tumor spatial information for the contrast-enhanced image in the training stage, which plays as guidance to determine the liver tumor location in the non-contrast-enhanced image. The results of the above two methods outperform the results of existing deep learning methods. The success of proposed methods in medical object detection and segmentation indicates the deep reinforcement learning can be a reliable computer-assisted-diagnosis tool and benefit to clinicians.
Zhang, Dong, "Deep Reinforcement Learning in Medical Object Detection and Segmentation" (2020). Electronic Thesis and Dissertation Repository. 7423.
Creative Commons License
This work is licensed under a Creative Commons Attribution-Noncommercial 4.0 License