本文已被：浏览 55次 下载 25次
Received:November 03, 2025 Published Online:May 01, 2026
Received:November 03, 2025 Published Online:May 01, 2026
中文摘要: 目的 探讨基于乳腺超声图像的深度学习可视化模型对低年资医师鉴别乳腺结节良恶性的辅助诊断价值。方法 回顾性收集2022年9月至2024年9月于上海市第六人民医院经超声筛查发现并经病理证实的乳腺结节病例（共420例）的临床资料及超声图像。按7∶3比例随机划分为训练集（n=294）与独立测试集（n=126）。基于训练集图像，构建并比较了VGG16、GoogleNet、AlexNet、ResNet50、Vision Transformer、DenseNet121六种端到端深度学习模型的性能，通过受试者工作特征（ROC）曲线下面积（AUC）、准确率、灵敏度、特异度、精确率、F1值（精确率和召回率的调和平均值）等指标筛选出最优模型并生成可视化结果。采用自身前后对照设计，由低年资医师在不知晓病理结果的情况下，独立对测试集图像进行诊断［基于乳腺影像报告与数据系统（BI-RADS）分类］；间隔2周后，同一低年资医师借助筛选出的最优可视化模型对相同图像进行辅助诊断。计算并比较低年资医师在独立诊断与模型辅助诊断模式下的诊断效能指标。结果 ResNet50模型在训练集上表现最优，其AUC为 0.937（95%CI ：0.911~0.962），准确率、灵敏度、特异度、精确率、F1 值分别为 0.864、0.863、0.865、0.863、0.863。该模型的核心诊断效能指标优于参与研究的超声医师。低年资超声医师在 ResNet50 可视化热图辅助下，二次阅片对乳腺结节诊断的AUC、准确率、灵敏度、特异度分别由首次的0.597、59.5%、80.8%、38.5%，提高至0.904、90.5%、84.9%和95.9%。结论 基于乳腺超声的深度学习可视化模型（如ResNet50）具有良好的乳腺结节良恶性鉴别能力。该模型能有效辅助低年资医师提高诊断敏感性及整体效能，可作为提升低年资医师乳腺超声诊断规范化水平、缩短培训周期的有价值的辅助工具。
Abstract:Objective To evaluate the diagnostic value of a deep learning visualization model based on breast ultrasound images in assisting junior sonographers to differentiate benign and malignant breast nodules. Methods The clinical data and ultrasound images from 420 pathologically confirmed breast nodule cases identified through screening at Shanghai Sixth People ??s Hospital from September 2022 to September 2024 were collected. Patients were randomly divided into a training set（n=294）and an independent test set（n=126）in a 7 to 3 ratio. Six end-to-end deep learningmodels（VGG16，GoogleNet，AlexNet，ResNet50，Vision Transformer，DenseNet121）were constructed and compared using the training set images. Performance was evaluated using the area under the receiver operating characteristic（ROC）curve（AUC），accuracy，sensitivity，specificity，precision，and F1 value（harmonic average of accuracy and recall）. The optimal model was selected and used to generate visualizations. Employing a self - controlled before - after design，a junior sonographer independently interpreted the test set images blinded to pathological results，assigning Breast Imaging Reporting and Data System（BI -RADS）categories. After a two -week washout period，the same junior sonographer re - interpreted the same images with the assistance of the ResNet50 visualization model. Diagnostic performance metrics were calculated and compared for the junior sonographer under independent and model - assisted reading conditions. Results The ResNet50 model demonstrated optimal performance on the training set，achieving an AUC of 0.937（95%CI ：0.911-0.962），with accuracy，sensitivity，specificity，precision，and F1 values of 0.864，0.863，0.865，0.863 and 0.863，respectively. The core diagnostic performance metrics of the model surpassed those of the participating sonographer. With the assistance of ResNet50 visualization heatmaps，the AUC，accuracy，sensitivity，and specificity of junior sonographer in the second reading of breast nodule diagnosis increased from the initial 0.597，59.5%，80.8%，38.5% to 0.904，90.5%，84.9%，and 95.9%，respectively. Conclusion Deep learning visualization models based on breast ultrasound images，such as ResNet50，exhibit strong capability in differentiating benign and malignant breast nodules. These models effectively assist junior sonographers by improving diagnostic sensitivity and overall performance，serving as valuable auxiliary tools to enhance standardized diagnostic skills and potentially shorten training cycles.
keywords: Breast nodules Ultrasonography Differential diagnosis Deep learning Human - machine comparison Visualization
文章编号： 中图分类号：R737.9 R445.1 文献标志码：A
基金项目:
引用文本：