目的 探讨基于超声的迁移学习人工智能模型预测甲状腺囊实性结节（PCTN）恶性概率的可能性。 方法 回顾性分析2021年1月至2023年12月间就诊于山东大学齐鲁医学院威海市立医院并有明确病理结果的PCTN患者246例，以7：3的比例随机分为训练组和测试组。评估PCTN超声图像特征，经过多因素logistic回归分析，得到评估PCTN恶性概率的独立危险因素并计算曲线下面积（AUC）。另一方面，通过Python软件的PyTorch框架对数据进行预处理后，选择5种不同的预训练模型进行迁移学习，具体包括Inception_v3、EfficientNet、VGG19、ResNet50和DenseNet121，计算AUC值并进行比较。 结果 超声图像特征中实性成分> 50%、实性成分与囊性成分呈偏心锐角、病灶边界模糊不清、病灶边界呈毛刺样、蛋壳样钙化和微钙化对于评价PCTN的良恶性差异有统计学意义（P < 0.05），基于以上独立危险因素计算的AUC值为0.843。另外，在5种迁移学习模型中，ResNet50模型诊断效能最高，AUC值为0.903 2。 结论 基于超声的迁移学习人工智能模型优于传统超声图像评价效能，能够准确预测PCTN的性质，从而减少不必要的超声引导下细针穿刺活检。

Objective To investigate the feasibility and accuracy of an ultrasound-based transfer learning artificial intelligence model in predicting the malignancy probability of partially cystic thyroid nodules （PCTN）. Methods A retrospective analysis was conducted on 246 patients with PCTN who had definitive pathological results and were admitted to Weihai Municipal Hospital， Cheeloo College of Medicine， Shandong University from January 2021 to December 2023. Patients were randomly divided into training and test cohorts at a ratio of 7：3. Ultrasonic image features of PCTN were evaluated， and independent risk factors were identified using multivariate logistic regression analysis， with the area under the curve （AUC） subsequently calculated. Additionally， five different pre-trained models?Inception_v3， EfficientNet， VGG19， ResNet50， and DenseNet121?were selected for transfer learning after data preprocessing using the PyTorch framework in Python. The AUC values of these models were calculated and compared. Results Solid portion greater than 50%， eccentric acute angle， ill-defined margin， spiculated or microlobulated margin， rim calcification， and microcalcification exhibited statistically significant differences（P < 0.05） in distinguishing between benign and malignant PCTN. The AUC value derived from these independent risk factors was 0.843. Furthermore， among the five transfer learning models evaluated， the ResNet50 model demonstrated the highest diagnostic efficiency， achieving an AUC value of 0.903 2. Conclusion The ultrasound-based transfer learning artificial intelligence model demonstrated superior performance compared to traditional ultrasound image evaluation methods， enabling accurate prediction of the nature of PCTN and thereby reducing unnecessary ultrasound-guided fine needle biopsies.