面向胶质母细胞瘤的自主组织毁损：深度学习分割与机器人控制整合
本研究针对胶质母细胞瘤术中边界划定困难的问题，提出一种人工智能与机器人结合的闭环框架。深度学习模型基于术中超声数据进行实时肿瘤分割，分割结果整合机器人控制以引导组织毁损靶向。在小鼠GBM模型的离体与在体实验中，靶向总体与预期治疗区域一致，仅有轻微欠射。
要点
- 术中超声的深度学习分割可实时输出肿瘤区域，供机器人引导靶向。
- 离体与在体实验中组织毁损靶向总体与预期区域一致，仅轻微欠射。
- 该框架为胶质母细胞瘤治疗自动化提供了临床前可行性证据。
核心结论
- 结合人工智能超声分割与机器人引导进行组织毁损靶向在临床前胶质母细胞瘤模型中具有可行性。
研究设计与样本
- 研究类型为人工智能与机器人整合的闭环框架开发及临床前评估。
- 样本为小鼠胶质母细胞瘤模型，实验包括离体与在体两部分。
- 深度学习模型基于术中超声数据训练，用于实时肿瘤分割。
关键影像表现/征象
- 离体与在体实验中组织毁损靶向总体与预期治疗区域一致，表现为轻微欠射。
诊断标准与关键数值
- 深度学习模型在术中超声数据上实现了较强的实时肿瘤分割性能。
临床应用与阅片要点
- 术中超声实时分割结果可作为机器人组织毁损靶向的依据。
- 靶向偏差主要呈欠射表现，阅片或操作时需关注治疗区覆盖范围。
局限性
- 原文未报告具体样本量、分割性能数值及欠射的定量程度。
- 研究基于临床前小鼠模型，尚未在人体中验证。
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: Glioblastoma multiforme (GBM) is an aggressive brain tumor in which incomplete margin delineation during surgery can contribute to residual disease or unintended damage to healthy tissue. This study proposes an artificial intelligence- and robotics-enabled closed-loop framework for automated tumor delineation and guided histotripsy ablation.
METHODS: Deep learning models were trained on intra-operative ultrasound data to perform real-time tumor segmentation. The segmentation output was integrated with robotic control to guide histotripsy targeting. The framework was evaluated in mouse GBM models using ex vivo and in vivo experiments.
RESULTS: The models achieved strong real-time tumor segmentation performance. In ex vivo and in vivo experiments, histotripsy targeting generally aligned with the intended treatment region, with minimal undershooting.
CONCLUSION: These findings support the feasibility of combining artificial intelligence-based ultrasound segmentation with robotic guidance for histotripsy targeting in preclinical GBM models, as a step toward increased treatment automation.
引用
- Dorosti, Landry, Croul, et al. Toward Autonomous Histotripsy: Integrating Deep Learning Segmentation With Robotic Control for Glioblastoma. Ultrasound Med Biol. 2026. PMID: 42225519. doi:10.1016/j.ultrasmedbio.2026.04.023
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。