검색 상세

A Study on Unsupervised Deep Neural Networks for Medical Ultrasound Imaging

  • 조현우 (Hyunwoo Cho, 서강대학교 일반대학원)

초록 (요약문)

Ultrasound imaging is a widely used diagnostic tool due to its real-time capabilities, non- invasive nature, and cost-effectiveness. However, the inherent presence of speckle noise, artifacts, and clutter can significantly impede the diagnostic quality of ultrasound images. While deep learning has emerged as a powerful tool to address these challenges, traditional supervised approaches often falter due to the difficulty in acquiring paired training datasets, a critical requirement for their success. This dissertation delves into the potential of unsupervised learning frameworks to circumvent these limitations and elevate ultrasound image quality across various domains. This research focuses on enhancing B- mode image reconstruction by leveraging the inherent signal coherence properties within ultrasound data. By emphasizing spatial and temporal coherence, the proposed methods achieve remarkable improvements in resolution and contrast without relying on ground- truth data. This adaptability to diverse imaging conditions and patient-specific scenarios offers a robust foundation for superior diagnostic performance, overcoming the limitations of conventional beamforming methods. Furthermore, this work tackles the pervasive issue of artifacts and clutter, which plague ultrasound images and hinder accurate diagnoses. By modeling artifact behaviors and employing adaptive loss weighting techniques, the proposed frameworks effectively suppress these unwanted signal components while preserving crucial anatomical details. Notably, these methods exploit angle-dependent variations in artifacts to achieve robust performance in both flow phantom and in vivo studies, showcasing their critical implications for applications like perfusion imaging and surpassing the generalizability of conventional model-based methods. The versatility of these unsupervised learning frameworks extends to diverse imaging modalities, encompassing both B-mode and perfusion imaging. By effectively suppressing clutter and preserving flow signals, these methods ensure accurate perfusion analysis and facilitate high-resolution anatomical and functional imaging. This adaptability across different imaging needs provides a comprehensive solution for modern ultrasound imaging, ensuring seamless application and enhanced diagnostic capabilities. Addressing the challenge of speckle noise, which degrades image quality and obscures critical details, this research also introduces innovative frameworks that effectively reduce speckle while preserving essential structural information. By leveraging unsupervised learning and exploiting angle-dependent speckle variations, these methods achieve superior performance compared to traditional approaches. This advancement translates to clearer, more diagnostically informative images, particularly valuable in applications where subtle details are crucial for accurate interpretation. In conclusion, this dissertation investigates the transformative potential of unsupervised learning in revolutionizing ultrasound imaging. By circumventing the reliance on paired training data and offering adaptability across various imaging domains, the proposed frameworks pave the way for scalable, efficient, and clinically impactful solutions. The research not only enhances image quality but also expands the possibilities of ultrasound as a diagnostic tool, promising significant advancements in healthcare and disease management.

more

목차

1. Chapter 1 Introduction 3
1.1. Medical ultrasound 3
1.2. Deep learning approaches for medical ultrasound 7
1.3. Challenges of deep learning driven medical ultrasound 10
1.3.1. Conventional supervised learning frameworks 10
1.3.2. Ill-posed problems of image reconstruction tasks 11
1.3.3. Data acquisition for conventional deep learning models 12
1.4. Specific aims and thesis organization References 12
References 14
2. Chapter 2 B-Mode Image Reconstruction 18
2.1. Introduction 18
2.2. Conventional ultrasound beamformers 22
2.3. Proposed framework: Deep coherence learning 25
2.3.1. Basic principles and problem formulations 25
2.3.2. Unsupervised learning framework 26
2.3.3. Network architecture and implementation details 29
2.4. Experimental setups 31
2.4.1. Dataset utilization 31
2.4.2. Evaluation metrics 34
2.5. Experimental results 35
2.5.1. Training curves 35
2.5.2. Simulation and phantom study 36
2.5.3. In vivo study 44
2.5.4. Computation complexity analysis 50
2.5.5. Ablation study: Impact of angle span 50
2.6. Discussion 52
2.7. Conclusion 55
References 56
3. Chapter 3 Reverberation Artifact Reduction 65
3.1. Introduction 65
3.2. Proposed unsupervised learning framework 68
3.2.1. Basic principles and overall framework 68
3.2.2. Modeling of reverberation artifact and dataset utilization 70
3.2.3. Network architecture and loss weighting techniques 72
3.3. Experimental setups 74
3.3.1. Implementation details and evaluation metrics 74
3.4. Experimental results 75
3.4.1. Phantom study: Fabricated artifact models 75
3.4.2. In vivo study 84
3.5. Discussion 86
3.6. Conclusion 88
References 89
4. Chapter 4 Clutter Rejection for Perfusion Imaging 92
4.1. Introduction 92
4.2. Conventional clutter filters for perfusion imaging 94
4.3. Proposed framework: Deep clutter filter 97
4.3.1. Basic principles and overall framework 97
4.3.2. Training and inference stages 99
4.3.3. Network architecture 103
4.4. Experimental setups 104
4.4.1. Dataset utilization and implementation 104
4.4.2. Evaluation metrics 107
4.5. Experimental results 107
4.5.1. Flow phantom study 108
4.5.2. In vivo study 113
4.6. Discussion 116
4.7. Conclusion 119
References 119
5. Chapter 5 Speckle Reduction and Image Enhancement 122
5.1. Introduction 122
5.2. Conventional speckle reduction methods 124
5.3. Proposed unsupervised learning frameworks 127
5.3.1. Basic principles and overall frameworks 127
5.3.2. Unsupervised speckle reduction technique 129
5.3.3. Self-supervised image enhancement technique 132
5.3.4. Network architecture 132
5.4. Experimental setups 134
5.4.1. Dataset utilization and implementation 134
5.4.2. Evaluation metrics 135
5.5. Experimental results 136
5.5.1. Simulation phantom study 136
5.5.2. Experimental phantom study 139
5.5.3. In vivo study 143
5.6. Discussion 150
5.7. Conclusion 151
References 152
6. Chapter 6 Conclusions and Future Directions 156
6.1. Introduction 156
6.2. Contributions 157
6.2.1. High quality B-mode image reconstruction 157
6.2.2. Artifact and clutter rejection 158
6.2.3. Applicable to various imaging modes 159
6.2.4. Speckle reduction and image enhancement 160
6.3. Future directions 161
6.3.1. Network compression and acceleration using hardware 161
6.3.2. Generalization and system integration 162
6.4. Conclusions 163

more
반출 Meta View 목록