Ultrasound localization microscopy using Deep Learning

Restricted (Penn State Only)
- Author:
- Liu, Xilun
- Graduate Program:
- Computer Science and Engineering
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- July 15, 2024
- Committee Members:
- Chitaranjan Das, Program Head/Chair
Mohamed Almekkawy, Chair & Dissertation Advisor
William Higgins, Major Field Member
Daniel Cortes Correales, Outside Unit & Field Member
Syed Rafiul Hussain, Major Field Member - Keywords:
- Ultrasound Localization Mircroscopy
Deep Learning
Microbubbles Localization
Super Resolution - Abstract:
- Ultrasound localization microscopy (ULM) is a super-resolution imaging technique that uses ultrasound waves and Microbubbles (MBs) to achieve a high spatial resolution in biological tissues. MBs are used as scatterers for the ultrasound waves. MBs are typically injected into the bloodstream and allowed to accumulate in the region of interest. The ultrasound waves then interact with the MBs, generating echoes that can be detected and used to localize the MBs with a high precision. The ULM employs an iterative algorithm to refine the positions of scatterers and reconstruct a high-resolution image of the sample. This technique can achieve resolutions beyond the diffraction limit, which allows visualization of structures and features that were previously too small to be resolved. The detection and localization of MBs are limited by the drawbacks of traditional methods, such as reduced accuracy and robustness and the potential for computational resource exhaustion. Moreover, traditional methods typically rely on prior knowledge for the manual selection of thresholds to achieve a high performance. This approach may lead to limitations in accuracy and robustness, which are common challenges in localization methods. Furthermore, the manual selection process can be time consuming and prone to errors, reducing the efficiency of the overall process. These limitations can lead to challenges in achieving precise and reliable localization of MBs, which is essential for the accurate diagnosis and monitoring of various medical conditions. The need for alternative approaches to overcome these limitations is crucial to improve the efficiency and effectiveness of MB localization. Deep neural networks (DNNs) have gained noticeable acceleration in development in the last decade, branching into applications in various professional and academic field, such as medical signals and imaging processing. DNNs have demonstrated the ability to extract abstract features from medical images and to generate predictive patterns. The availability of graphics processing units has significantly reduced the computation time required to solve complex problems in medical image analysis using DNNs. Consequently, DNN-based approaches have become promising tools for medical image interpretation and diagnosis, offering the potential for improved accuracy and efficiency in clinical practice. The ultimate objective of this work is to introduce an accurate and computationally efficient localization method for ULM that has the potential to enable rapid and precise imaging of biological tissues at high resolution. In this thesis, several deep learning models are introduced to optimize the localization step. First, a Swin transformer-based neural network is proposed to perform end-to-end mapping to implement MBs localization. The effectiveness of the proposed method was validated using synthetic and in vivo data with various quantitative metrics. To further improve the processing efficiency and data memory management, object detection models were designed. This approach has been extended to 3D applications. The performance of the proposed method was evaluated by comparing it with the traditional approaches in terms of accuracy and efficiency. The results demonstrate that the proposed networks achieve a more precise and better imaging capability than previously used methods. The resolution of the ULM using our proposed method on the in vivo rat data can reach 29 \(\mu m\) in 2D and 125 \(\mu m\) in 3D. In addition, the proposed method offer a significant reduction in the computational cost per frame compared to traditional methods, with processing times as low as 0.003 s, even in the 3D ULM. These improvements in the computational efficiency make this technique feasible for time-sensitive applications.