面向伽玛刀放射外科治疗前计划剂量学可行性的物理感知几何深度学习

文献信息

中文摘要

在伽玛刀立体定向放射外科（SRS）中，实现最佳剂量学结果受到复杂肿瘤几何形状的物理约束。虽然梯度指数（GI）和适形指数（CI）定义了计划质量，但接受[公式：见文本]的体积和靶点密度决定患者安全性和效率。当前的体积启发式方法无法捕捉支配这些参数的拓扑不规则性。本研究引入一种物理感知的几何深度学习框架，直接从表面拓扑同时预测质量、安全性和复杂性指标。方法：利用Vestibular-Schwannoma-SEG数据集[公式：见文本]，将肿瘤流形转换为物理感知点云，并富集12维特征向量，包括表面法线和径向曲率。开发了一种基于PointNet的多任务架构，称为“GammaPoint-Net”，用于回归GI、CI、[公式：见文本]和靶点密度。该框架与基于3D体素的CNN、标准体积回归、在手工3D形状描述符上训练的传统机器学习回归器（随机森林、支持向量回归和XGBoost）以及基于点的架构进行了基准比较，采用5折交叉验证。GammaPoint-Net在所有目标上均取得更优的预测准确性（[公式：见文本]），显著优于体积基线[公式：见文本]。消融研究证实，纳入微分几何特征使平均绝对误差降低46%。GammaPoint-Net也优于最强的手工形状描述符基线（增强曲率矩的XGBoost，[公式：见文本]；配对Wilcoxon[公式：见文本]）。显著性映射表明，网络自主识别高曲率表面突起是剂量学劣化的主要驱动因素。GammaPoint-Net证明，表面拓扑而非体积是放射外科计划质量的基本决定因素。该框架提供即时、可解释的治疗前计划评估，使物理师能够在优化前预判计划难度和毒性风险。

英文摘要

In Gamma Knife stereotactic radiosurgery (SRS), achieving optimal dosimetric outcomes is physically constrained by complex tumor geometry. While the Gradient Index (GI) and Conformity Index (CI) define plan quality, the volume receiving [Formula: see text] and Shot Density dictate patient safety and efficiency. Current volumetric heuristics fail to capture the topological irregularities that govern these parameters. This study introduces a physics-aware Geometric Deep Learning framework to simultaneously predict quality, safety, and complexity metrics directly from surface topology. Methods: Utilizing the Vestibular-Schwannoma-SEG dataset [Formula: see text], tumor manifolds were converted into physics-aware point clouds enriched with 12-dimensional feature vectors, including surface normals and radial curvature. A multi-task PointNet-based architecture, termed "GammaPoint-Net," was developed to regress GI, CI, [Formula: see text], and Shot Density. The framework was benchmarked against 3D Voxel-based CNNs, standard volumetric regression, conventional machine-learning regressors trained on handcrafted 3D shape descriptors (Random Forest, Support Vector Regression, and XGBoost), and point-based architectures using 5-fold cross-validation. GammaPoint-Net achieved superior predictive accuracy across all objectives ([Formula: see text]), significantly outperforming volumetric baselines [Formula: see text]. Ablation studies confirmed that incorporating differential geometric features reduced the mean absolute error by 46%. GammaPoint-Net also outperformed the strongest handcrafted shape-descriptor baseline (XGBoost augmented with curvature moments, [Formula: see text]; paired Wilcoxon [Formula: see text] ). Saliency mapping demonstrated that the network autonomously identifies high-curvature surface protrusions as the primary drivers of dosimetric degradation. GammaPoint-Net demonstrates that surface topology, rather than volume, is the fundamental determinant of radiosurgical plan quality. This framework provides an instantaneous, interpretable pre-planning assessment, enabling physicists to anticipate plan difficulty and toxicity risks prior to optimization.