Unsupervised Domain Adaptation for Deep Learning based Tissue Speed-of-Sound Estimation in Medical Ultrasound Imaging
- 주제어 (키워드) Medical ultrasound , Speed-of-sound estimation , Quantitative imaging , Deep learning , Unsupervised Domain adaptation
- 발행기관 서강대학교 일반대학원
- 지도교수 유양모
- 발행년도 2024
- 학위수여년월 2024. 2
- 학위명 석사
- 학과 및 전공 일반대학원 전자공학과
- 실제URI http://www.dcollection.net/handler/sogang/000000077018
- UCI I804:11029-000000077018
- 본문언어 영어
- 저작권 서강대학교 논문은 저작권 보호를 받습니다.
초록
Medical ultrasound is one of the most preferred imaging modalities due to its clinical usefulness such as non-ionization, real-time capability. Traditionally, B-mode imaging is widely used for diagnosis of anatomical structures and tissues. However, B- mode imaging only offers limited tissue characteristics while other characteristics such as tissue elasticity can also provide clinically important information. Recent studies have proposed quantitative ultrasound imaging such as shear wave elastography, attenuation imaging and these quantitative imaging techniques showed promising results for differentiate suspicious tissues (e.g., malignant breast lesions). Quantitative imaging often requires special hardware systems or complex computations, as well as high framerate ability. These limitations make the quantitative imaging difficult to adopt on various circumstances such as portable devices. Recently, deep learning-based methods have been proposed for speed-of-sound estimation of ultrasound from raw radio-frequency signals and showed great potential for high framerate speed-of-sound imaging. However, the main challenge of these deep learning-based speed-of-sound imaging is the dataset acquisition. Since the accurate speed-of-sound map on real experimental data is extremely difficult to acquire, the most conventional studies used simulation dataset. The ultrasound simulation cannot reflect complex physics of real circumstances which may cause a domain discrepancy which degrades the performance of deep learning model on real circumstances. In a very recent study, to mitigate the domain discrepancy of deep learning-based speed-of-sound imaging, a pre-processing technique extracting phase information was used to remove the gain-dependency. However, the limitations regarding the probe and system characteristics remain. For the domain discrepancy problem, domain adaptation can mitigate the discrepancy. When the ground-truth of target domain data is available, transfer learning or supervised domain adaptation techniques can be used. However, for speed-of-sound imaging, the ground-truth of target domain (real data) in not available which requires unsupervised domain adaptation. Such unsupervised domain adaptation techniques need to be designed with certain prior knowledges and unsupervised domain adaptation for speed-of-sound imaging is not introduced yet. Therefore, in this paper, unsupervised domain adaptation techniques are used to enhance the robustness of the deep learning-based speed-of-sound imaging by mitigating the domain discrepancy between the simulated data and real data. The robustness of the introduced method is evaluated on simulation data and ex-vivo data. The introduced method can be potentially used for various clinical applications with improved reliability.
more목차
1 Introduction 1
2 Methods and Materials 4
2.1 Overall framework 4
2.2 Network architecture and simulated source domain dataset 5
2.3 Unsupervised domain adaptation for speed-of-sound imaging 7
2.3.1 Distance metric learning 8
2.3.2 Self-training 10
2.4 Uncertainty calibration for pseudo-labeling of SoS 11
2.4.1 Channel-wise uncertainty calibration 12
2.4.2 Angle-wise uncertainty calibration 13
2.5 Network Training 14
2.6 Experimental setup and evaluation metrics 15
2.6.1 Experimental setup 15
2.6.2 Evaluation metrics 17
3. Results 18
3.1 Simulation study 18
3.2 Ex-vivo study 20
4. Discussions 22
5. Conclusions 23
6. References 24
