构建颈动脉斑块超声图像分类数据集，构建一种甄别颈动脉斑块稳定性的双输入BCNN-ResNet深度学习模型，探讨双输入BCNN-ResNet模型对颈动脉斑块稳定性自动分类及诊断的效能。

收集2021年1月至2023年3月在上海市第八人民医院、大连大学附属新华医院接受颈动脉超声检查者493例。由4名超声科医师观察颈动脉斑块声像图，经综合评判后选取颈动脉稳定斑块超声图像352张，易损斑块超声图像691张，构建共包含1043张颈动脉超声图像的数据集。使用ResNet-50模型作为基础模型，第一个ResNet-50网络输入斑块结构图像，提取结构特征；第二个ResNet-50网络输入裁剪后的斑块图像，获取像素特征，融合两组特征构建一个双输入BCNN模型。通过对图像进行分类监督学习的训练、内部验证和外部验证，比较新构建双输入BCNN-ResNet-50模型与ResNet-34、ResNet-50、ResNet-101、单输入BCNN-ResNet-34、双输入BCNN-ResNet-34、单输入BCNN-ResNet-50模型对颈动脉斑块稳定性的分类诊断效能。应用ROC曲线下面积、敏感度、特异度、准确性、真阳性、假阳性等指标评估模型的诊断效能。

ROC曲线结果显示，双输入BCNN-ResNet-50模型对超声颈动脉斑块稳定性分类诊断的曲线下面积（AUC）为0.896。单输入BCNN-ResNet-50模型AUC为0.878，而ResNet-34、ResNet-50、ResNet-101、单输入BCNN-ResNet-34、双输入BCNN-ResNet-34模型AUC分别为0.857、0.860、0.859、0.864、0.868。双输入BCNN-ResNet-50模型对于颈动脉斑块的稳定性数据集分类及诊断的效能明显优于其他模型。

双输入BCNN-ResNet模型可以自动甄别超声颈动脉斑块稳定性，此算法优于以往诊断模型，为后续临床颈动脉斑块稳定性筛查应用提供了技术参考。

To construct a carotid plaque ultrasound image dataset, build a dual-input BCNN-ResNet classification deep learning model for screening carotid plaque stability, and explore the efficacy of the dual-input bilinear convolutional neural network with residual network as the backbone network (BCNN-ResNet) model for automatic classification and diagnosis of carotid plaque stability.

A total of 493 cases were collected from January 2021 to March 2023 from those who underwent carotid ultrasonography at the Shanghai Eighth People's Hospital and Xinhua Hospital of Dalian University. Carotid plaque images were observed by four ultrasonographers, and 352 ultrasound images of stable carotid plaques and 691 images of vulnerable plaques were selected after comprehensive evaluation to construct a dataset containing a total of 1043 ultrasound images of carotid arteries. Using the ResNet-50 model as the base model, with the first ResNet-50 network inputting the structural plaque images to extract the structural features and the second ResNet-50 network inputting the cropped plaque images to obtain the pixel features, a dual-input BCNN model was constructed by fusing the two sets of features. By training the images with classification supervised learning, internal validation, and external validation, the newly constructed dual-input BCNN-ResNet model was compared with ResNet-34, ResNet-50, ResNet-101, single-input BCNN-ResNet-34, dual-input BCNN-ResNet-34, and single-input BCNN-ResNet-50 for diagnostic efficacy in classifying carotid plaque stability. The diagnostic efficacy of the model was assessed by applying metrics such as the area under the receiver operating characteristic (ROC) curve (AUC).

ROC curve analysis showed that the AUC of the dual-input BCNN-ResNet-50 model for the classification and diagnosis of carotid plaque stability on ultrasound images was 0.896, while the AUC values of the single-input BCNN-ResNet-50 model, ResNet-34, ResNet-50, ResNet-101, single input BCNN-ResNet-34, and dual-input BCNN-ResNet-34 model AUC were 0.878, 0.857, 0.860, 0.859, 0.864, and 0.868, respectively. The efficacy of the dual-input BCNN-ResNet-50 model was significantly better than that of the other models in classifying and diagnosing carotid plaque stability.

The dual-input BCNN-ResNet model can automatically screen ultrasound images for carotid plaque stability, and this algorithm outperforms previous diagnostic models, providing a technical reference for clinical carotid plaque stability screening.