本研究旨在基于超声影像组学特征、深度学习特征及临床特征，构建术前预测甲状腺乳头状癌（PTC）患者中央区颈部淋巴结转移（CLNM）的融合模型，并评价其性能。

回顾性分析了2020年12月至2023年3月期间在复旦大学附属肿瘤医院接受手术治疗并经病理证实的510例PTC患者。所有患者术前均接受超声引导下细针穿刺活检。术后病理结果显示，387例患者CLNM为阴性，123例患者为阳性。由经验丰富的超声医师在灰阶超声图像上手工勾画病灶感兴趣区域。采用Pyradiomics提取影像组学特征，利用预训练的 VGG16-BN 模型提取深度学习特征。通过特征筛选与降维获得最终影像组学与深度学习特征，并结合临床特征构建融合特征集。所有病例按4∶1比例划分为训练集与独立测试集，训练阶段采用五折交叉验证进行模型构建与参数优化，最终基于支持向量机（SVM）建立预测模型。模型性能在独立测试集中评估，并计算ROC曲线下面积（AUC）及其95%置信区间。

ROC曲线结果显示，仅影像组学模型的判别能力最弱（AUC=0.63），仅深度学习模型性能显著提升（AUC=0.91），深度学习联合影像组学的双模态融合模型进一步提高（AUC=0.92），全特征融合模型（影像组学+深度学习+临床特征）取得最优交叉验证性能（AUC=0.93）。在独立测试集中，三模态融合模型的AUC为0.962（95%置信区间：0.915~0.994），特异度为 0.987（95%置信区间：0.931~0.999），敏感度为0.583（95%置信区间：0.366~0.779），显示出较好的整体判别能力及较高的阴性排除能力。

融合超声影像组学、深度学习特征及临床危险因素构建的多模态预测模型能够有效提高PTC患者CLNM的术前风险评估能力，具有潜在的临床辅助决策价值，其泛化性能仍需在多中心、大样本研究中进一步验证。

To develop a preoperative predictive model for central compartment lymph node metastasis (CLNM) in papillary thyroid carcinoma (PTC) by integrating ultrasound radiomics features, deep learning features, and clinical characteristics and to evaluate its performance.

A retrospective analysis was conducted on 510 pathologically confirmed PTC patients who underwent surgical treatment at Fudan University Shanghai Cancer Center between December 2020 and March 2023. All patients received ultrasound-guided fine-needle aspiration biopsy preoperatively. Postoperative pathology identified 387 CLNM-negative and 123 CLNM-positive cases. Regions of interest were manually delineated on grayscale ultrasound images by experienced radiologists. Radiomics features were extracted using Pyradiomics, and deep learning features were obtained via a pretrained VGG16-BN model. After feature selection and dimensionality reduction, the final radiomics and deep learning features were combined with clinical features to construct a multimodal feature set. All cases were split into training and independent test sets at a 4∶1 ratio. Five-fold cross-validation was applied during training for model construction and hyperparameter optimization, and a support vector machine (SVM)-based predictive model was established. Model performance was evaluated on the independent test set using the area under the receiver operating characteristic curve (AUC) and its 95% confidence interval (CI).

ROC curve analysis showed that the discriminative ability of the radiomics-only model was the weakest (AUC=0.63), while the performance of the deep learning-only model was significantly improved (AUC=0.91). The bimodal fusion model combining deep learning and radiomics further enhanced the performance (AUC=0.92), and the full-feature fusion model (radiomics + deep learning + clinical features) achieved the best cross-validation performance (AUC=0.93). On the independent test set, the trimodal fusion model achieved an AUC of 0.962 (95%CI: 0.915-0.994), a specificity of 0.987 (95%CI: 0.931-0.999), and a sensitivity of 0.583 (95%CI: 0.366-0.779), demonstrating strong overall discriminative ability and high negative exclusion capability.

The multimodal predictive model integrating ultrasound radiomics, deep learning features, and clinical risk factors effectively enhances preoperative risk stratification for CLNM in PTC patients, offering potential clinical decision-support value. Its generalizability warrants further validation in multicenter, large-sample studies.