解剖学先验信息神经网络：在损失函数与网络架构中编码解剖学先验，并结合SE(3)公式建立导丝诱导的腹主动脉髂动脉变形模型
文章背景与核心概要
应用於生物结构的深度学习模型往往会生成在数值上合理但在解剖学上不可能实现的输出，导致在训练数据稀缺时泛化能力较差。本文介绍了解剖学先验信息神经网络（Anatomy-Informed Neural Networks, AINN），这是一个旨在将临床知识直接嵌入深度学习模型的新颖框架。
该框架通过在损失函数中引入惩罚项来实现软解剖学先验（例如，惩罚从髂动脉而非主动脉分出的异常肾移植动脉），并将硬解剖学先验直接嵌入网络架构和状态表示中，从而在结构上彻底杜绝产生无效预测的可能性。这种方法在一个受限的临床测试案例中得到了验证：预测在腔内引入硬导丝时腹主动脉髂动脉树如何发生变形——这是现代主动脉手术和自主血管内导航的关键问题。
📋 摘要 (Summary)
Deep-learning models applied to biological structures often generate outputs that are numerically plausible yet anatomically impossible, resulting in poor generalization when training data is scarce. This paper introduces Anatomy-Informed Neural Networks (AINN), a novel framework designed to embed clinical knowledge directly into deep learning models:
- Soft Anatomic Priors: Integrated as penalty terms within the loss function (e.g., penalizing an unexpected renal transplant artery off the iliac instead of the aorta).
- Hard Anatomic Priors: Embedded directly into the network architecture and state representation to render invalid predictions impossible by construction.
The methodology is demonstrated on a constrained clinical test case: predicting how the aortoiliac tree deforms during the endoluminal introduction of a stiff guidewire—a critical problem for modern aortic surgery and autonomous endovascular navigation.
应用於生物结构的深度学习模型往往会生成在数值上合理但在解剖学上不可能实现的输出，导致在训练数据稀缺时泛化能力较差。本文介绍了解剖学先验信息神经网络（Anatomy-Informed Neural Networks, AINN），这是一个旨在将临床知识直接嵌入深度学习模型的新颖框架：
- 软解剖学先验（Soft Anatomic Priors）： 作为惩罚项集成到损失函数中（例如，惩罚从髂动脉而非主动脉分出的异常肾移植动脉）。
- 硬解剖学先验（Hard Anatomic Priors）： 直接嵌入网络架构和状态表示中，从结构上使得无效预测不可能发生。
该方法在一个受限的临床测试案例中得到了验证：预测在腔内引入硬导丝时腹主动脉髂动脉树如何发生变形——这是现代主动脉手术和自主血管内导航的关键问题。
🔬 方法论与框架 (Methodology & Framework)
- State Representation: The vessel centerline and wire path are lifted from \(\mathbb{R}^3\) to curves of frames within the Lie group \(SE(3)\).
- Mechanical Modeling: A Cosserat-rod wire is coupled to a tortuosity-modulated, anatomically anchored vessel via a unilateral lumen-contact inequality. The resulting prediction is a constrained minimizer of the coupled elastic energy, utilizing contact forces as its Lagrange multipliers.
- Supervision: Employs a Wasserstein-2 optimal-transport loss comparing the predicted projection (through the C-arm geometry) against the observed angiogram, allowing 2D angiograms to effectively supervise 3D predictions.
- 状态表示： 血管中心线和导丝路径从 \(\mathbb{R}^3\) 提升（lift）到李群 \(SE(3)\) 中的坐标系曲线。
- 力学建模： 采用Cosserat杆导丝，通过单侧管腔接触不等式将其与受弯曲度调节、具有解剖学锚定的血管耦合。由此产生的预测结果是耦合弹性能量的约束极小化器，并利用接触力作为其拉格朗日乘子。
- 监督机制： 采用Wasserstein-2最优传输损失，将预测投影（通过C形臂几何结构）与观测到的血管造影图进行比较，从而允许二维血管造影图有效地监督三维预测。
📊 评估与当前状态 (Evaluation & Current Status)
- Verification: The kinematics, loss functions, and spatial projections have been successfully verified against known ground truth. The mechanics solver is validated against its own optimality conditions (though predicted displacements are not yet mesh-converged).
- Training Status: No neural network has been trained in this preliminary work. Future research will scale this in silico model to real CT scans to evaluate whether it enhances predictive accuracy while reducing the volume of required training data.
- 验证： 运动学、损失函数和空间投影已针对已知的真实值（ground truth）进行了成功验证。力学求解器根据其自身的优化条件进行了验证（尽管预测的位移尚未达到网格收敛）。
- 训练状态： 在这项初步工作中尚未训练任何神经网络。未来的研究将把这一计算机模拟（in silico）模型扩展到真实的CT扫描数据，以评估它是否能在减少所需训练数据量的同时提高预测准确性。
🔗 全文与访问链接 (Full-Text & Access Links)
(License icon source preserved as required: 
)