Motion Estimation in Ultrasound Image Sequences: A Deep Learning Approach
Open Access
- Author:
- Bharadwaj, Skanda
- Graduate Program:
- Computer Science and Engineering
- Degree:
- Master of Science
- Document Type:
- Master Thesis
- Date of Defense:
- March 17, 2021
- Committee Members:
- Mohamed Khaled Almekkawy, Thesis Advisor/Co-Advisor
William Evan Higgins, Committee Member
Robert Collins, Committee Member
Chitaranjan Das, Program Head/Chair - Keywords:
- motion estimation
object tracking
ultrasound imaging
deep learning
Siamese networks
convolutional neural networks
Kalman Filter - Abstract:
- Motion estimation in ultrasound images using speckle tracking has been known to be very beneficial for image-guided radiation therapy. However, the conventional exhaustive search-based block matching technique is computationally expensive and infeasible for real-time processing. In this study, we first explore the conventional block matching-based similarity matching using faster search algorithms and then introduce a deep learning-based similarity learning technique for robust real-time landmark tracking. In lieu of the exhaustive search strategy, a faster search algorithm called the Adaptive Rood Pattern Search (ARPS) is adopted to track regions of interest in ultrasound images along with sub-pixel matching to reduce the computational cost and enhance block matching. This experiment proved that faster search algorithms come at the cost of accuracy. Thus we explore much more sophisticated techniques to reduce the computational complexity without compromising the accuracy. Towards this end, we explore similarity learning using deep learning techniques. In this methodology, we adopt the Fully-Convolutional Siamese Network (SiamFC) to track regions of interest. Despite the remarkable efficiency of SiamFC, it suffers two major limitations. First, it considers a constant position model and, therefore, is extremely sensitive to the deformation of the reference object. Second, it is a detection-based tracker and does not consider any motion model. Consequently, tracking the reference object is solely dependent on the detections obtained in each frame. In order to overcome these limitations, an upgraded version of the SiamFC is proposed in this technique. The two limitations of SiamFC are overcome with the introduction of two modules into the original architecture: 1) A template update module with frame anchoring is introduced into SiamFC making it robust to deformation of the reference object resolving the issue of constant position model, and 2) A linear/non-linear motion model is introduced based on the tissue under consideration that addresses the second issue of the original SiamFC. Kalman Filter (Linear Kalman Filter for linear motion model and Extended Kalman Filter for non-linear motion model) is designed to incorporate the motion model into the proposed architecture. The experiments provided promising results towards robust landmark tracking in ultrasound images for image-guided radiation therapy.