Humbert-Vidan Laia, Hansen Christian R, Patel Vinod, Johansen Jørgen, King Andrew P, Guerrero Urbano Teresa
Department of Medical Physics, Guy's and St Thomas' NHS Foundation Trust, London, UK.
School of Cancer and Pharmaceutical Sciences, Comprehensive Cancer Centre, King's College London, London, UK.
Phys Imaging Radiat Oncol. 2024 Nov 2;32:100668. doi: 10.1016/j.phro.2024.100668. eCollection 2024 Oct.
While the inclusion of spatial dose information in deep learning (DL)-based normal-tissue complication probability (NTCP) models has been the focus of recent research studies, external validation is still lacking. This study aimed to externally validate a DL-based NTCP model for mandibular osteoradionecrosis (ORN) trained on 3D radiation dose distribution maps and clinical variables.
A 3D DenseNet-40 convolutional neural network (3D-mDN40) was trained on clinical and radiation dose distribution maps on a retrospective class-balanced matched cohort of 184 subjects. A second model (3D-DN40) was trained on dose maps only and both DL models were compared to a logistic regression (LR) model trained on DVH metrics and clinical variables. All models were externally validated by means of their discriminative ability and calibration on an independent dataset of 82 subjects.
No significant difference in performance was observed between models. In internal validation, these exhibited similar Brier scores around 0.2, Log Loss values of 0.6-0.7 and ROC AUC values around 0.7 (internal) and 0.6 (external). Differences in clinical variable distributions and their effect sizes were observed between internal and external cohorts, such as smoking status (0.6 vs. 0.1) and chemotherapy (0.1 vs. -0.5), respectively.
To our knowledge, this is the first study to externally validate a multimodality DL-based ORN NTCP model. Utilising mandible dose distribution maps, these models show promise for enhancing spatial risk assessment and guiding dental and oncological decision-making, though further research is essential to address overfitting and domain shift for reliable clinical use.
虽然将空间剂量信息纳入基于深度学习（DL）的正常组织并发症概率（NTCP）模型一直是近期研究的重点，但仍缺乏外部验证。本研究旨在对基于DL的下颌骨放射性骨坏死（ORN）NTCP模型进行外部验证，该模型是在三维辐射剂量分布图和临床变量上进行训练的。
在一个由184名受试者组成的回顾性类平衡匹配队列的临床和辐射剂量分布图上训练了一个三维密集连接网络40卷积神经网络（3D-mDN40）。第二个模型（3D-DN40）仅在剂量图上进行训练，并将这两个DL模型与在剂量体积直方图（DVH）指标和临床变量上训练的逻辑回归（LR）模型进行比较。所有模型均通过其判别能力以及在一个由82名受试者组成的独立数据集上的校准进行外部验证。
各模型之间未观察到性能上的显著差异。在内部验证中，这些模型的布里尔分数均在0.2左右，对数损失值在0.6 - 0.7之间，受试者工作特征曲线下面积（ROC AUC）值在内部约为0.7，在外部约为0.6。在内部和外部队列之间观察到临床变量分布及其效应大小的差异，例如吸烟状况（分别为0.6对0.1）和化疗（分别为0.1对 -0.5）。
据我们所知，这是第一项对基于多模态DL的ORN NTCP模型进行外部验证的研究。利用下颌骨剂量分布图，这些模型在增强空间风险评估和指导牙科及肿瘤学决策方面显示出前景，不过要解决过拟合和域转移问题以实现可靠的临床应用，还必须进行进一步研究。
