被動空化成像可用於監測高強度聚焦超音波治療所產生的微氣泡振盪和破裂，它通常是通過對聲學陣列被動接收的數據進行波束成形來構建空化事件的空間分佈，但傳統被動波束成形的空間分辨率較差而難以準確定位空化的具體發生位置。提高被動空化成像分辨率的方法之一是直接根據接收到的通道波形的時間延遲分佈來預測空化位置，這可以通過相鄰通道之間的波形互相關計算來估計，然而互相關計算的準確性會隨噪聲位準提高而降低。
在這項研究中我們將接收到的通道數據視為圖像輸入U-Net++深度卷積神經網絡來估計時間延遲分佈，U-Net++結構與傳統U-Net相似但增加嵌套和密集跳躍連接來提高延遲分佈的估計精度。使用來自感興趣區域內各種空化位置的模擬通道數據和相應的真實延遲曲線作為訓練數據，訓練過的U-Net++ 可描繪從單個氣泡發出的通道數據的延遲曲線，之後再搭配擬合延遲曲線即可找到氣泡位置。本研究還開發了預處理算法以考慮兩個同時聲源之通道波形可能互相重疊的情形。
本研究以模擬與實驗方式驗證所提出的被動空化成像方法。單氣泡模擬結果顯示U-Net++ 神經網絡估計的RF 通道延遲曲線在通道SNR 為 -10 dB下可達到空化位置估計誤差為1.22 mm，此時傳統互相關預測的氣泡位置會明顯偏離真實值。雙氣泡情形的位置估計在兩個空化通道數據沒有重疊的情況下能夠準確的估計兩個空化信號的位置，但是在通道數據出現高度重疊的情況下可能出現預測誤差較大或是完全無法預測的情形。
整體而言，本研究所提出的方法可以在單氣泡或低度重疊的多氣泡情形下穩定地提供氣泡位置的準確預測以及利用預測的位置進行空化影像的重建，相較傳統波束成像算法有著更快的被動空化成像速度和較好的空間解析度。

Passive cavitation imaging can be used to monitor the therapy of high-intensity focused ultrasound to avoid overtreatment. It is conventionally constructed by beamforming the data passively received by an acoustic array to image the spatial distribution of cavitation events due to the generation, oscillation and collapse of microbubbles. Nonetheless, the poor spatial resolution of passive beamforming makes it difficult to accurately determine the specific location of cavitation. An alternative approach is to predict the cavitation site directly from the time delay profile of the received channel waveform which can be estimated by correlation between adjacent channels. The accuracy of correlation method, however, degrades in the presence of high noises.
In this study, a deep convolutional neural network (U-Net++) is proposed to estimate the time delay profile by regarding the received channel data as an image input. U-Net++ is based on nested and dense skip connections to enhance the estimation accuracy of delay profile. Using the simulated channel data from various cavitation sites within the region of interest and the corresponding true delay profiles as the training data, U-Net++ has learned to predict the delay curve of channel data emitted from a single bubble. The delay curve is then fitted to find the bubble position. Pre-processing algorithm is also developed to consider possible overlapping of channel waveforms from two simultaneous sources.
Both simulations and experiments are performed to verify the proposed method. For the single-bubble case, the predicted bubble position in the correlation method markedly deviates from the true value as the SNR decreases. However, the proposed method can still predict the correct cavitation position with an error of 1.22 mm when the channel SNR decreases to -10 dB. For the double-bubble case, the two cavitation positions can be accurately estimated when their channel signals do not overlap. In the presence of overlap of channel signals, however, the proposed method may suffer from larger estimation error.

[1] 沈哲州，「醫用超音波上課講義」，國立台灣科技大學電機所，民國107年。Che-Chou Shen, and Pei-Ying Hsieh.
[2] Y. Wang , Z. B. Wang , and Y. H. Xu . " Efficacy Efficiency, and Safety of Magnetic Resonance Guided High - Intensity Focused Ultrasound for Ablation of Uterine Fibroids : Comparison with Ultrasound - Guided Method " Korean Journal of Radiology Article vol . 19 , no . 4 , pp . 724-732 , Jul - Aug 2018.
[3] H. Azzouz and J. J. M. C. H. de la Rosette , " HIFU : Local Treatment of Prostate Cancer, EAU - EBU Update Series , vol . 4, no. 2. pp. 62-70., (2006).
[4] Elisa E Konofagou 1, Yao-Sheng Tung, James Choi, Thomas Deffieux, Babak Baseri, Fotios Vlachos (2012). Ultrasound-induced blood-brain barrier opening. Current pharmaceutical biotechnology, 13(7), 1332-1345.
[5] De Senneville BD, Moonen C, Ries M. MRI-Guided HIFU Methods for the Ablation of Liver and Renal Cancers. Adv Exp Med Biol. 2016;880:43-63. doi: 10.1007/978-3-319-22536-4_3. PMID: 26486331.
[6] W.S. Chen, A.A. Brayman, T.J. Matula, L.A. Crum, M.W. Miller The pulse length dependence of inertial cavitation dose and hemolysis Ultrasound Med. Biol., 29 (2003), pp. 739-748.
[7] S. J. Norton and I. J. Won, "Time exposure acoustics," in IEEE Transactions on Geoscience and Remote Sensing, vol. 38, no. 3, pp. 1337-1343, May 2000, doi: 10.1109/36.843027.
[8] X. Ma et al., "Multiple Delay and Sum With Enveloping Beamforming Algorithm for Photoacoustic Imaging," in IEEE Transactions on Medical Imaging, vol. 39, no. 6, pp. 1812-1821, June 2020, doi: 10.1109/TMI.2019.2958838.
[9] S. Lu, H. Hu, X. Yu, J. Long, B. Jing, Y. Zong, M. WanPassive acoustic mapping of cavitation using eigenspace-based robust Capon beamformer in ultrasound therapy Ultrason. Sonochem., 41 (2018), pp. 670-679, 10.1016/j.ultsonch.2017.10.017.
[10] C. Coviello et al., “Passive acoustic mapping utilizing optimal beamforming in ultrasound therapy monitoring,” J. Acoust. Soc. Amer., vol. 137, no. 5, pp. 2573–2585, May 2015.
[11] Lu S, Li R, Yu X, Wang D, Wan M. Delay multiply and sum beamforming method applied to enhance linear-array passive acoustic mapping of ultrasound cavitation. Med Phys. 2019 Oct;46(10):4441-4454. doi: 10.1002/mp.13714. Epub 2019 Aug 10. PMID: 31309568.
[12] P . Boulos, F .Varray, A .Poizat, A .Ramalli, B. Gilles, JC. Bera, C. Cachard. Weighting the Passive Acoustic Mapping Technique With the Phase Coherence Factor for Passive Ultrasound Imaging of Ultrasound-Induced Cavitation. IEEE Trans Ultrason Ferroelectr Freq Control.2018Dec;65(12):2301-2310.doi: 10.1109/TUFFC.2018.2871983. Epub 2018 Sep 24. PMID: 30273149.
[13] A.V. Telichko, T. Lee, M. Jakovljevic and J. J. Dahl, "Passive Cavitation Mapping by Cavitation Source Localization From Aperture-Domain Signals—Part I: Theory and Validation Through Simulations," in IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control, vol. 68, no. 4, pp. 1184-1197, April 2021, doi: 10.1109/TUFFC.2020.3035696.
[14] A.V. Telichko, T. Lee, M. Jakovljevic and J. J. Dahl, "Passive Cavitation Mapping by Cavitation Source Localization From Aperture-Domain Signals—Part I: Theory and Validation Through Simulations," in IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control, vol. 68, no. 4, pp. 1184-1197, April 2021, doi: 10.1109/TUFFC.2020.3035696. M. E. Anderson and G. E. Trahey, “The direct estimation of sound speed using pulse-echo ultrasound,” Citation, The J. Acoust. Soc. Amer., vol. 104, p. 3099, 1998.
[15] M. E. Anderson and G. E. Trahey, “The direct estimation of sound speed using pulse-echo ultrasound,” Citation, The J. Acoust. Soc. Amer., vol. 104, p. 3099, 1998.
[16] Y. Liu et al., "Artifact Suppression for Passive Cavitation Imaging Using U-Net CNNs with Uncertainty Quantification," 2019 IEEE 4th International Conference on Signal and Image Processing (ICSIP), 2019, pp. 1037-1042, doi: 10.1109/SIPROCESS.2019.8868593.
[17] H. Sharma, et al., "Deep convolutional neural networks for automatic classification of gastric carcinoma using whole slide images in digital histopathology," Computerized Medical Imaging and Graphics 61:2-13, 2017.
[18] E. Garcia, et al., "Automatic lymphocyte detection on gastric biomedical image segmentation," Proc. MICCAI, pp. 234-241, 2015.
[19] O. Ronnebergerm et al., "U-net: Convolutional networks for biomedical image segmentation," Proc. MICCAI, pp. 234-241, 2015.
[20] J. Long, E. Shelhamer, and T. Darrell. Fully convolutional networks for semantic segmentation. arXiv preprint arXiv:1411.4038, 2014. 1, 2, 4, 6, 7.
[21] K. J. Haworth, K. B. Bader, K. T. Rich, C. K. Holland and T. D. Mast, "Quantitative Frequency-Domain Passive Cavitation Imaging," in IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control, vol. 64, no. 1, pp. 177-191, Jan. 2017, doi: 10.1109/TUFFC.2016.2620492.
[22] Juejin.cn. 2022. 指纹特征提取及描述（附Python代码） - 掘金. [online] Available at: <https://juejin.cn/post/6924305350497828872> [Accessed 13 September 2022].
[23] Burgess MT, Apostolakis I, Konofagou EE. Power cavitation-guided blood-brain barrier opening with focused ultrasound and microbubbles. Phys Med Biol. 2018 Mar 15;63(6):065009. doi: 10.1088/1361-6560/aab05c. PMID: 29457587; PMCID: PMC5881390.