视觉Transformer算子网络加速异质性脊髓解剖中的聚焦超声建模
本研究开发了基于视觉Transformer的DeepONet(ViT-DeepONet),用于实时预测脊髓聚焦超声压力场,以克服传统波传播模拟耗时数小时的问题。模型在8000张猪脊髓模拟压力图上训练,测试损失3.1%,泛化到人超声图像损失4.4%,计算时间较CPU高保真求解器减少逾91000倍,且性能优于CNN-DeepONet(损失5.1%)而参数少五倍。结果表明算子学习可加速声学建模并优化患者特异性治疗超声成像。
要点
- ViT-DeepONet可实现脊髓聚焦超声压力场的实时预测,测试损失3.1%。
- 模型泛化到人超声图像损失4.4%,计算时间减少逾91000倍。
- ViT-DeepONet优于CNN-DeepONet(损失5.1%)且参数少五倍。
核心结论
- ViT-DeepONet能准确捕捉异质性声学域中的空间与几何依赖,实现低计算负担的实时超声压力预测。
- 算子学习可作为加速声学建模和优化患者特异性治疗超声成像的有效框架。
研究设计与样本
- 研究为数据驱动框架开发,训练数据为8000张来自25头猪脊髓的模拟压力图。
- 模型为基于视觉Transformer的DeepONet(ViT-DeepONet),并与卷积DeepONet(CNN-DeepONet)及几何感知傅里叶神经算子进行基准比较。
- 通过学习控制偏微分方程的解算子,模型无需新模拟即可泛化到不同换能器位置和脊髓几何。
诊断标准与关键数值
- ViT-DeepONet测试损失3.1%,泛化到人超声图像损失4.4%。
- CNN-DeepONet损失5.1%,ViT-DeepONet参数少五倍。
- 几何感知傅里叶神经算子损失最低(2.6%),但复杂度高八倍。
- 计算时间较基于CPU的高保真求解器减少逾91000倍。
临床应用与阅片要点
- 该技术有望用于聚焦超声治疗脊髓损伤前的实时压力场预测,辅助换能器定位决策。
- 当前证据基于模拟数据和猪脊髓,尚未在临床实时工作流中验证,阅片时需注意其研究阶段。
局限性
- 训练数据来自猪脊髓模拟压力图(n=25),样本量有限且为临床前数据。
- 研究未报告在人体临床环境中的前瞻性验证或实际手术决策中的效用。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
英文原文(摘要)· 点击展开
OBJECTIVE: The purpose of this work is to develop a data-driven framework for real-time prediction of focused ultrasound pressure fields in the spinal cord, addressing the limitations of computationally intensive acoustic simulations. Focused ultrasound therapy offers submillimeter precision for targeted treatment of spinal cord injuries, but its efficacy depends critically on transducer placement due to the spinal cord's complex geometry and acoustic heterogeneity across anatomical layers.
METHODS: Conventional computational models rely on wave propagation simulations to generate patient-specific pressure maps, but these can take hours to complete parameter sweeps and are impractical for surgical decision-making. To address this bottleneck, we introduce a vision transformer-based Deep Operator Network (ViT-DeepONet), benchmarked against a convolutional DeepONet (convolutional neural network [CNN]-DeepONet) and a geometry-aware Fourier Neural Operator. By learning the solution operator of the governing partial differential equations, these models generalize across transducer positions and spinal geometries without new simulations.
RESULTS: Trained on 8000 simulated pressure maps from porcine spinal cords (n=25), ViT-DeepONet achieves realtime predictions with 3.1% test loss and generalizes to human ultrasound images with 4.4% loss, reducing computation time by more than 91,000-fold relative to central processing unit-based high-fidelity solvers. ViT-DeepONet outperforms CNN-DeepONet (5.1% loss) while using fivefold fewer parameters. A geometry-aware Fourier Neural Operator attains the lowest loss (2.6%), but with eightfold higher complexity.
CONCLUSION: ViT-DeepONet accurately captures spatial and geometric dependencies in heterogeneous acoustic domains, enabling real-time ultrasound pressure prediction with a lower computational burden. These results demonstrate the feasibility of operator learning as a powerful framework for accelerating acoustic modeling and optimizing patient-specific therapeutic ultrasound imaging.
引用
- Kumar A, Zhi X, Davidar AD, et al. Accelerating Focused Ultrasound Modeling in Heterogeneous Spinal Cord Anatomy With Vision Transformer Operator Networks. Ultrasound Med Biol. 2026. PMID: 42580951. doi:10.1016/j.ultrasmedbio.2026.07.014
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。