儿童医院如何用开源 NVIDIA AI 做心脏护理

费城儿童医院正在使用开源人工智能工具，在几秒钟内模拟儿童的心脏——目标是为先天性心脏病儿童提供更安全、更精确的护理。

大约1%的活产儿存在先天性心脏缺陷——而且没有两个缺陷是完全相同的。

一个孩子如果出生时心脏下腔之间有孔，或单个泵血腔的瓣膜漏血而维持生命，就需要符合其具体解剖结构的护理。历史上，外科医生使用的装置几乎从未以该特定儿童为设计目标。

“你拥有一个独一无二的孩子，却是一个现成的装置，”费城儿童医院（CHOP）的心脏病学家兼研究员Matthew Jolley博士说。“我们的工作是找到合适的——而建模让我们可以在任何人进入导管室或手术室之前完成这一步。”

CHOP 的心脏建模服务建立在 MONAI 之上：https://project-monai.github.io/ — 一个由 NVIDIA 共同创立的开源医学影像框架 — 它使用儿童护理团队已拥有的图像，如 CT 扫描、MRI 和三维超声，并在几秒钟内生成解剖精确的心脏模型。曾经需要熟练研究人员四小时完成的工作流程，现在快到足以用于常规临床。

美国已有20多家儿童医院运行心脏建模项目。在波士顿儿童医院，建模支持超过一半的心脏手术——大约每年500例手术。CHOP预计今年将完成约200例建模案例。从心脏护理领域开始的这项工作，CHOP现在计划通过其摩根研究与创新中心下的IDEA实验室：https://www.research.chop.edu/building-breakthroughs-morgan-centers-idea-lab-invites-innovation-with-elizabeth-silvestro-phd，在多个学科中应用相同的工具。

当Jolley在2015年加入CHOP时，3D超声心动图刚刚开始使用。虽然已有用于建模成人瓣膜的工具，但几乎没有为他所治疗的复杂小型解剖结构而设计的工具。

他的实验室与开源社区合作，开发了SlicerHeart——3D Slicer开源软件的一个扩展，用于可视化、分割和分析3D医学图像——并开始开发工作流程，从多种影像模态中建模儿童的心脏和心脏瓣膜。

多年来，生成一个模型意味着一个熟练的研究助理需要在工作站上花费数小时。机器学习改变了这一点。

使用MONAI Label和NVIDIA的Auto3DSeg实现，Jolley的团队在先前影像和模型的配对上训练分割网络。输出达到受过训练的人工所能产生的同等质量——而且是几秒钟完成，而不是数小时。

“机器学习已经成为面包和黄油（基础工具）了，”Jolley说。“一旦我们制作了10或20对影像-模型，我们就训练一个模型并开始应用它。”

临床影响迅速显现。对于复杂的室间隔缺损——心脏下腔之间的孔洞——CHOP现在在手术前常规进行建模。

一个早期案例清楚地显示了其价值：一个孩子已经经历了两次失败的修复尝试，外科医生无法通过传统方法定位缺损。3D模型澄清了解剖结构。第一次尝试就成功完成了修复。

对于这样的病例，Jolley说，心脏建模已经从研究阶段转向标准护理流程。

仅仅可视化并不是最终目标。Jolley的团队不仅想知道孩子的心脏长什么样，还想知道在任何手术开始之前，当装置部署在其中时会发生什么。这就是Newton：https://developer.nvidia.com/newton-physics 的用武之地。

Newton是一个基于NVIDIA Warp：https://developer.nvidia.com/warp-python Python框架构建的开源物理引擎，该框架在GPU上运行物理模拟。CHOP正在与NVIDIA以及开源社区合作，使用Warp构建以生物力学为重点的模拟框架，并将其引入原本用于基于模拟的AI机器人训练的Newton。

这些框架，一旦与 3D Slicer 和 SlicerHeart 集成，就能帮助医生了解决定设备在特定患者中部署方式的组织材料特性。借助 GPU 加速，它们可以将心脏设备模拟所需的时间从最长四小时——或多种配置的整夜运行——缩短到接近实时。

在实际操作中，临床医生可以比较不同设备如何适合儿童的具体解剖结构，并快速获得结果，从而支持同一天的决策。

CHOP 已开始为用于封闭儿童心脏缺损的闭合装置实施基于 Warp 和 Newton 的功能——并希望将类似的方法应用于经导管瓣膜的模拟。Warp 和 Newton 的开源架构正在与 SlicerHeart 连接，长期目标是将实时模拟引入临床工作流程。

基于 SlicerHeart 和 NVIDIA Omniverse 的耦合器也在开发中：https://www.nvidia.com/en-us/omniverse/，使用 OpenUSD 提供支持：https://www.nvidia.com/en-us/glossary/openusd/。OpenUSD 的开源 3D 互操作性允许将各种数据和解算器整合到基于患者图像构建的模拟中。然后，这些模拟可以流入虚拟现实环境，并利用嵌入的视觉-语言模型（VLMs）：https://www.nvidia.com/en-us/glossary/vision-language-models/，让临床医生在执行操作前能够直观地查询和与儿童的心脏解剖模拟进行交互。

美国约有 240 万人患有先天性心脏病。历史上，该人群既稀少又多样化，无法吸引传统医疗设备公司按家庭所需的规模进行投资。没有任何一家公司的工具能够满足 Jolley 团队的需求，也没有单一机构能够独自构建这些工具。

SlicerHeart 的工具可免费使用和开发。斯坦福大学和波士顿儿童医院的研究人员在 CHOP 之外贡献了额外工具。一个由全国儿童医院组成的联盟正在形成，以构建下一代共享建模基础设施，并以开源作为各机构之间的连接纽带。

“通过正常的经济学，人口太少，无法支撑传统的商业开发，”Jolley说。“但这是一个如此重要的问题，以至于在研究界和慈善界的推动下，人们正在支持它。开源通过允许协作和进步而不受障碍，挑战了小型和异质性人口的传统经济学。”

NVIDIA 在开发开放平台方面的投资和与他人积极合作——用于医疗影像AI的 MONAI、用于物理模拟的 Newton、用于3D互操作性和虚拟现实的 OpenUSD——让像 Jolley 团队这样的团队可以使用在工业规模上维护的基础设施。儿童医院的实验室现在可以快速利用那些通常需要公司级工程团队来构建和维护的工具，同时受益于广泛的开源社区持续不断的创新。

了解 MONAI 是如何推进医疗影像AI 的：https://project-monai.github.io/

Children’s Hospital of Philadelphia is using open source AI tools to model children’s hearts in seconds — with the goal of enabling safer, more precise care for kids with congenital heart disease.

About 1% of all live births involve a congenital heart defect — and no two are alike.

A child born with a hole between the lower chambers of their heart, or a leaking valve in the single pumping chamber keeping them alive, needs care that fits their exact anatomy. Historically, the devices surgeons reach for were almost never designed with that specific child in mind.

“You’ve got a one-of-a-kind kid and an off-the-shelf device,” said Dr. Matthew Jolley, a cardiologist and researcher at Children’s Hospital of Philadelphia, or CHOP. “Our job is to find what fits — and modeling lets us do that before anyone goes into the cath lab or operating room.”

CHOP’s cardiac modeling service, built on MONAI ：https://project-monai.github.io/ — an open source medical imaging framework cofounded by NVIDIA — takes the images a child’s care team already has, such as CT scans, MRI and 3D ultrasound, and produces anatomically precise heart models in just seconds. A workflow that once required four hours of work by a skilled researcher now completes fast enough for routine clinical use.

More than 20 children’s hospitals across the U.S. now run cardiac modeling programs. At Boston Children’s Hospital, modeling supports more than half of all cardiac surgeries — roughly 500 cases a year. CHOP expects to reach about 200 modeled cases this year. Where this work began in cardiac care, CHOP is now aiming to apply the same tools across multiple disciplines through the IDEA Lab ：https://www.research.chop.edu/building-breakthroughs-morgan-centers-idea-lab-invites-innovation-with-elizabeth-silvestro-phd , part of the hospital’s Morgan Center for Research and Innovation.

When Jolley joined CHOP in 2015, 3D echocardiography was just coming online. There were tools for modeling adult valves, but almost nothing built for the complex, small anatomies he was treating.

His lab worked with the open source community to build SlicerHeart — an extension of 3D Slicer open source software for visualizing, segmenting and analyzing 3D medical images — and began developing workflows to model pediatric hearts and valves from multiple imaging modalities.

For years, producing a single model meant a skilled research assistant spending hours at a workstation. Machine learning changed that.

Using MONAI Label and NVIDIA’s Auto3DSeg implementation, Jolley’s team trained segmentation networks on pairs of prior images and models. The output meets the same quality standard a trained human would produce — in seconds rather than hours.

“Machine learning has become just bread and butter,” Jolley said. “As soon as we’ve made 10 or 20 image-model pairs, we train a model and start applying it.”

The clinical impact was seen quickly. For complex ventricular septal defects — holes between the heart’s lower chambers — CHOP now models routinely before surgery.

One early case made the value clear: A child had already undergone two failed repair attempts, with surgeons unable to locate the defect with traditional methods. The 3D model clarified the anatomy. The repair succeeded on the first try.

For cases like these, Jolley said, cardiac modeling has moved from research to standard of care.

Visualization alone isn’t the end goal. Jolley’s team wants to know not just what a child’s heart looks like, but what will happen when a device is deployed inside it — before any procedure begins. That’s where Newton ：https://developer.nvidia.com/newton-physics comes in.

Newton is an open source physics engine built on the NVIDIA Warp ：https://developer.nvidia.com/warp-python Python framework that runs physics simulations on GPUs. CHOP is working with NVIDIA and the open source community to build biomechanics-focused simulation frameworks with Warp that can be brought into Newton, originally intended for simulation-based AI robot training.

These frameworks, once integrated with 3D Slicer and SlicerHeart, can help doctors understand tissue material properties that determine how a device will deploy in a given patient. With GPU acceleration, they can reduce the time needed for cardiac device simulation from up to four hours — or a full overnight run for multiple configurations — to near real time.

In practice, a clinician could compare how different devices fit a child’s specific anatomy and get results fast enough to inform a same-day decision.

CHOP has begun to implement features built on Warp and Newton for the closure devices used to seal holes in children’s hearts — and hopes to apply similar methods for simulations of transcatheter valves. The open source architecture of Warp and Newton is being connected with SlicerHeart, with the long-term goal of bringing real-time simulations into clinical workflows.

A coupler using SlicerHeart and NVIDIA Omniverse ：https://www.nvidia.com/en-us/omniverse/ digital twins, powered by OpenUSD, ：https://www.nvidia.com/en-us/glossary/openusd/ is also in development. OpenUSD’s open source 3D interoperability allows diverse data and solvers to be integrated into simulations built from patient images. These simulations can then flow into virtual reality environments and harness embedded vision-language models (VLMs) ：https://www.nvidia.com/en-us/glossary/vision-language-models/ , allowing clinicians to intuitively query and interact with a child’s cardiac anatomy in simulation before acting on it.

About 2.4 million people in the U.S. live with congenital heart disease. Historically, this has been a population too rare and too diverse to attract traditional device-company investment at the scale families need. No single company has built the tools Jolley’s team requires. No single institution could build them alone.

SlicerHeart’s tools are free to use and build on. Researchers at Stanford and Boston Children’s contribute additional tools alongside CHOP. A national consortium of children’s hospitals is now forming to build the next generation of shared modeling infrastructure, with open source as the connective tissue across institutions.

“It’s too small a population to support traditional commercial development by normal economics,” Jolley said. “But it’s such an important problem that between the research community and philanthropy, people are getting behind it. Open source defies traditional economics for small and heterogenous populations by allowing collaboration and progress without barriers.”

NVIDIA’s investment and active collaboration with others in developing open platforms — MONAI for medical imaging AI, Newton for physics simulation, OpenUSD for 3D interoperability and virtual reality — gives teams like Jolley’s access to infrastructure maintained at industrial scale. A lab at a children’s hospital can now quickly harness tools that would otherwise require a company-sized engineering team to build and sustain, while benefiting from the continued innovation of a broad open source community.

Learn how MONAI ：https://project-monai.github.io/ is advancing medical imaging AI.