【摘要】 该研究旨在构建一个用于颈动脉斑块超声图像的有无判定的数据集，由1 165例受检者的1 761张超声图像组成。研究采用了一种融合了双线性卷积神经网络与残差神经网络的深度学习架构，即单输入BCNN-ResNet模型，以辅助临床医生通过颈动脉超声图像进行斑块的诊断。该模型经过训练以及内部和外部验证后，在内部验证中，ROC AUC达到了0.99，其95%置信区间为(0.91, 0.84)，在外部验证中ROC AUC为0.95，其95%置信区间为(0.96, 0.94)，此表现优于ResNet-34网络模型在内部验证中0.98 AUC的95%置信区间(0.99,0.95)和外部验证中0.94 AUC的95%置信区间(0.95, 0.92)。因此，单输入BCNN-ResNet网络模型展示了优异的诊断性能，为颈动脉斑块的自动识别提供了一种创新的解决方案。更多还原

【Abstract】 This study aims at developing a dataset for determining the presence of carotid artery plaques in ultrasound images, composed of 1 761 ultrasound images from 1 165 participants. A deep learning architecture that combines bilinear convolutional neural networks with residual neural networks, known as the single-input BCNN-Res Net model, was utilized to aid clinical doctors in diagnosing plaques using carotid ultrasound images. Following training, internal validation, and external validation, the model yielded an ROC AUC of 0.99(95% confidence interval: 0.91 to 0.84) in internal validation and 0.95(95%confidence interval: 0.96 to 0.94) in external validation, surpassing the Res Net-34 network model, which achieved an AUC of 0.98(95% confidence interval: 0.99 to 0.95) in internal validation and 0.94(95%confidence interval: 0.95 to 0.92) in external validation. Consequently, the single-input BCNN-Res Net network model has shown remarkable diagnostic capabilities and offers an innovative solution for the automatic detection of carotid artery plaques.更多还原