Inception-SE attention-driven bandwidth enhancement and denoising for radiation-induced acoustic signals(PDF)
《中国医学物理学杂志》[ISSN:1005-202X/CN:44-1351/R]
- Issue:
- 2026年第3期
- Page:
- 321-329
- Research Field:
- 医学信号处理与医学仪器
- Publishing date:
Info
- Title:
- Inception-SE attention-driven bandwidth enhancement and denoising for radiation-induced acoustic signals
- Author(s):
-
WANG Xinyi1; HE Yadi1; ZHAO Xinxin1; CHEN Boyong1; SONG Ting1; LI Yongbao2
- 1. School of Biomedical Engineering, Southern Medical University, Guangzhou 510515, China 2. Collaborative Innovation Center for Cancer Medicine/State Key Laboratory of Oncology in South China/Sun Yat-sen University Cancer Center, Guangzhou 510060, China
- Keywords:
- Keywords: radiation-induced acoustic imaging noise removal bandwidth recovery deep learning attention mechanism
- PACS:
- R318
- DOI:
- DOI:10.3969/j.issn.1005-202X.2026.03.007
- Abstract:
-
Abstract: Objective To propose deep learning-based bandwidth enhancement and denoising method for addressing the issues of bandwidth limitation and noise interference in radiation-induced acoustic (RA) signals utilized for radiotherapy dose monitoring, thereby improving the quality of RA signals and ultimately enhancing the accuracy of dose reconstruction based on these signals. Methods The k-Wave toolbox was employed to acoustically simulate the generation and propagation of RA signals, with bandwidth truncation and multiple noise components incorporated in the simulation to approximate realistic measurement conditions. Using the simulated bandwidth-limited noisy signals as network inputs, the proposed RAD-Att-Net model based on an Inception structure and a squeeze-and-excitation attention mechanism was trained to realize joint optimization of bandwidth recovery and noise suppression. Results Experimental results demonstrated that the signals restored by RAD-Att-Net exhibited high consistency with the ideal radiofrequency signals. Compared with the bandwidth-limited noisy signals, the restored signals had a reduction in mean squared error from 41.793±13.463 to 0.007±0.003, an increase in structural similarity index from 0.008±0.003 to 0.846±0.019, and an improvement in peak signal-to-noise ratio from (-6.776±1.625) dB to (31.213±1.435) dB. Furthermore, dose distributions reconstructed from different signals were evaluated using gamma analysis to assess the consistency between the reconstructed dose from the restored signals and that from the ideal radiofrequency signals. For the 3%/2 mm criterion, the gamma pass rates under the threshold conditions of Dose>0%, >30%, >50%, and >70% were 99.71%±0.29%, 98.15%±2.57%, 97.08%±3.99%, and 96.44%±5.98%, respectively. Conclusion The RAD-Att-Net model can effectively recover the bandwidth of RA signals and suppress noise, significantly improving signal quality and enhancing the accuracy of dose distribution reconstruction using the restored signals.
Last Update: 2026-03-30