Detail View
Physics-Informed Machine Learning for Acoustic Hologram Design in Precise Holographic Transcranial Ultrasound Neuromodulation
WEB OF SCIENCE
SCOPUS
- Title
- Physics-Informed Machine Learning for Acoustic Hologram Design in Precise Holographic Transcranial Ultrasound Neuromodulation
- DGIST Authors
- Moon Hwan Lee ; Jae Youn Hwang ; Jin Ho Chang
- Advisor
- 황재윤
- Co-Advisor(s)
- Jin Ho Chang
- Issued Date
- 2026
- Awarded Date
- 2026-02-01
- Type
- Thesis
- Description
- Acoustic hologram, Physics-informed machine learning, Transcranial focused ultrasound
- Abstract
-
Transcranial focused ultrasound (tFUS) has emerged as a promising neuromodulation technique because it is noninvasive and can modulate deep brain circuits with millimeter-scale spatial precision. However, conventional approaches typically rely on single-focus stimulation limiting stimulation flexibility, and skull-induced acoustic aberrations further hinders precision targeting. Acoustic holograms have been proposed as a flexible ultrasound stimulation tool, as they can freely shape focal patterns into desired forms. Current hologram design methods are largely based on traditional algorithms such as Gerchberg-Saxton (GS) algorithm or time reversal, which suffer from limited accuracy and high computational complexity, making adaptive use challenging. Therefore, for adaptive holographic transcranial ultrasound neuromodulation, there is a strong need to develop more efficient and accurate hologram design algorithms.
더보기
In this thesis, we propose a novel physics-informed machine learning framework for generating optimal acoustic holograms. Machine learning provides a powerful approach for solving nonconvex and irregular optimization problems, however, existing learning-based hologram design algorithms often adopt overly simplified assumptions about hologram implementation and surrounding acoustic environments, leading to discrepancies between simulation and experiments. By integrating a more accurate physical model into the framework and devise a new type of physics-based acoustic hologram, our framework enables hologram design that reflects realistic acoustic conditions, thereby improving both reconstruction accuracy and experimental reproducibility. The proposed framework is for holographic transcranial ultrasound neuromodulation but also has the potential to be extended to next-generation ultrasound applications, including targeted drug delivery, 3D bioprinting, and soft-matter manipulation.|경두개 집속 초음파(transcranial focused ultrasound)는 비침습적이며 밀리미터 수준의 공간 정밀도로 뇌 심부 회로를 조절할 수 있는 유망한 신경조절 기술로 주목받고 있다. 그러나 기존에는 단일 초점 자극을 사용하고 두개골에 의한 음향 왜곡이 발생하여 자극의 유연성과 정밀도가 제한된다. 음향 홀로그램은 원하는 형태의 초점 패턴을 자유롭게 형성할 수 있어 유연한 초음파 자극 도구로써 제안되었다. 현재의 음향 홀로그램 생성 방식은 주로 Gerchberg-Saxtion (GS) 알고리즘 혹은 시간 반전과 같은 전통적인 기법에 기반하고 있어 정확도에 한계가 있으며 계산 속도가 느리거나 계산 복잡도가 높아 실시간 적용이 어렵다. 따라서 실시간 홀로그래픽 경두개 초음파 신경조절 적용을 위해 보다 효율적이고 정확한 음향 홀로그램 생성 알고리즘의 개발이 요구된다.
본 연구에서는 새로운 물리정보 머신러닝 프레임워크를 기반으로 최적의 음향 홀로그램을 생성하는 방법을 제안한다. 머신러닝은 비볼록 및 비정형 문제를 해결하는데 강력한 접근법을 제공하지만 기존의 머신러닝 기반 홀로그램 생성 알고리즘은 홀로그램 구현과 주변 음향 환경에 대해 지나치게 단순화된 가정을 두어, 시뮬레이션과 실험 간에 오차가 발생할 수 있다. 본 연구에서는 보다 정밀한 물리 모델을 학습 과정에 통합하고 새로운 물리 기반 음향 홀로그램 구현 방식을 제안함으로써, 실제 음향 물리 조건을 반영한 홀로그램 생성을 가능하게 하고 재구성 정확도와 실험 재현성을 동시에 향상시켰다. 제안된 프레임워크는 홀로그래픽 경두개 신경 조절 뿐만 아니라 정밀 약물 전달, 3D 바이오프린팅, 소프트 물질 제어 등 차세대 초음파 응용 기술로 확장될 수 있는 잠재력을 지닌다.
- Table Of Contents
-
1. Introduction 1
2. Background 9
2.1 Principles and Challenges of Transcranial Focused Ultrasound (tFUS) 9
2.2 Acoustic Hologram for transcranial holographic ultrasound (tHUS) 12
2.3 Computational Modeling and Acoustic Field Simulation 14
2.4 Physics-Informed Machine Learning 18
2.5 Remaining Challenges and Research Needs 20
3. Deep learning-based framework for fast and accurate acoustic hologram generation 23
3.1 Introduction 23
3.2 Materials and Methods 26
3.2.1 Environment for Acoustic Holography 26
3.2.2 Proposed Framework 28
3.2.3 Evaluation through Simulations and Experiments 35
3.3 Results 38
3.3.1 Numerical Evaluations 38
3.3.2 More Numerical Evaluations: Ablation Study 41
3.3.3 Experimental Evaluations 49
3.4 Discussion 50
3.5 Conclusion 54
4. Differentiable Physics-Informed Optimization of 3D-Printed Acoustic Holograms in Heterogeneous Media 55
4.1 Introduction 55
4.2 Materials and Methods 58
4.2.1 Thickness-Only Acoustic Hologram Optimization Framework 58
4.2.2 Application: Acoustic field reconstruction through a heterogeneous skull medium 68
4.3 Results 75
4.3.1 TOH optimization results 75
4.3.2 In-Silico Performance 77
4.3.3 Ex-Vivo Field Validation 86
4.4 Discussion 88
4.5 Conclusion 92
5. Physics-Informed Self-Supervised Learning Framework for Holographic Transcranial Ultrasound Using a 2D Matrix Array 93
5.1 Introduction 93
5.2 Materials and Methods 96
5.2.1 Problem Formulation and Framework Overview 96
5.2.2 Differentiable Acoustic Propagation Model 98
5.2.3 Network Architecture and Training 100
5.2.4 Dataset and Experimental Setup 106
5.3 Results 111
5.3.1 Implementation Details 111
5.3.2 Skull-aware Focusing Accuracy 111
5.3.3 Performance across Target Complexity 113
5.3.4 Summary of Performance Comparison 115
5.4 Discussion 116
5.5 Conclusion 119
6. Summary and Conclusions 120
Appendix A 124
A.1 Stereotaxic Coordinates of the Target Regions 124
A.2 Acoustic Material Properties and Fabrication Details 124
A.3 Custom designed and fabricated ultrasound transducer module 124
A.4 Detailed Experimental Setup for Ex-Vivo Acoustic Field Measurements 126
A.5 Calibration Procedure for Projection-Based Measurement 126
A.6 Hydrophone Scanning Parameters and Data Acquisition Protocol 127
A.7 Reconstruction of 3D Fields via Angular Spectrum Method 128
A.8 Thermal simulations 128
A.9 Effect of Target Focal Volume Diameter 129
BIBLIOGRAPHY 131
- URI
-
https://scholar.dgist.ac.kr/handle/20.500.11750/59628
http://dgist.dcollection.net/common/orgView/200000948132
- Degree
- Doctor
- Publisher
- DGIST
File Downloads
- There are no files associated with this item.
공유
Total Views & Downloads
???jsp.display-item.statistics.view???: , ???jsp.display-item.statistics.download???:
