头颈癌 (HNC) 患者的传统正常组织并发症概率 (NTCP) 模型通常基于单值变量，对于辐射引起的口干症，该模型是基线口干症和平均唾液腺剂量。本研究旨在利用来自辐射剂量分布、CT 成像、危险器官分割和深度学习 (DL) 临床变量的 3D 信息来改进晚期口干症的预测。来自两个研究所的 1208 名 HNC 患者的国际队列研究用于训练和两次验证 DL 模型（DCNN、EfficientNet-v2 和 ResNet），以 3D 剂量分布、CT 扫描、危险器官分割、基线口干评分、性别和年龄作为输入。 NTCP 终点为放疗后 12 个月的中度至重度口干症。将深度学习模型的预测性能与参考模型进行比较：最近发布的口干症 NTCP 模型，使用基线口干症评分和平均唾液腺剂量作为输入。创建注意力图是为了可视化 DL 预测的焦点区域。进行迁移学习以提高深度学习模型在外部验证集上的性能。在独立测试中，所有基于深度学习的 NTCP 模型均表现出比参考 NTCP 模型（AUCtest=0.74）更好的性能（AUCtest=0.78 - 0.79）。注意力图显示，深度学习模型重点关注主要唾液腺，特别是腮腺干细胞丰富的区域。深度学习模型获得的外部验证性能（AUCexternal=0.63）低于参考模型（AUCexternal=0.66）。在小型外部子集上进行迁移学习后，DL 模型（AUCtl，external=0.66）的表现优于参考模型（AUCtl，external=0.64）。当在来自同一个研究所。通过迁移学习提高了外部数据集的性能，证明了需要多中心训练数据来实现基于深度学习的通用 NTCP 模型。版权所有 © 2024。由 Elsevier Inc. 发布。

Conventional normal tissue complication probability (NTCP) models for head and neck cancer (HNC) patients are typically based on single-value variables, which for radiation-induced xerostomia are baseline xerostomia and mean salivary gland doses. This study aims to improve the prediction of late xerostomia by utilizing 3D information from radiation dose distributions, CT imaging, organ-at-risk segmentations, and clinical variables with deep learning (DL).An international cohort of 1208 HNC patients from two institutes was used to train and twice validate DL models (DCNN, EfficientNet-v2, and ResNet) with 3D dose distribution, CT scan, organ-at-risk segmentations, baseline xerostomia score, sex, and age as input. The NTCP endpoint was moderate-to-severe xerostomia 12 months post-radiotherapy. The DL models' prediction performance was compared to a reference model: a recently published xerostomia NTCP model that used baseline xerostomia score and mean salivary gland doses as input. Attention maps were created to visualize the focus regions of the DL predictions. Transfer learning was conducted to improve the DL model performance on the external validation set.All DL-based NTCP models showed better performance (AUCtest=0.78 - 0.79) than the reference NTCP model (AUCtest=0.74) in the independent test. Attention maps showed that the DL model focused on the major salivary glands, particularly the stem cell-rich region of the parotid glands. DL models obtained lower external validation performance (AUCexternal=0.63) than the reference model (AUCexternal=0.66). After transfer learning on a small external subset, the DL model (AUCtl, external=0.66) performed better than the reference model (AUCtl, external=0.64).DL-based NTCP models performed better than the reference model when validated in data from the same institute. Improved performance in the external dataset was achieved with transfer learning, demonstrating the need for multicenter training data to realize generalizable DL-based NTCP models.Copyright © 2024. Published by Elsevier Inc.