摘要
目的本研究旨在利用深度学习中的卷积神经网络(convolutional neural networks,CNN)技术,对超声和胸部CT图像中的良性和恶性乳腺肿块进行鉴别。方法我们回顾性地收集了281张良性肿块图像和155张恶性肿块图像,所有病例均经过手术切除和麦默通活检以获得病理结果。采用ResNet-50的CNN架构,并集成了卷积块注意力模块(convolutional block attention module,CBAM),分别构建了超声(Ultrasound,US)、CT和US+CT的深度学习模型,对这些病变的良恶性进行了分析。我们计算了模型的灵敏度、特异性、准确度以及受试者工作特征曲线下面积(Area Under the Curve,AUC),并与人工阅片结果进行了比较。结果超声科医生阅读超声图像得到的AUC为0.808,而CT医生阅读CT图像得到的AUC为0.749。单独的超声模型的AUC为0.883,单独的CT模型的AUC为0.871,而超声加CT联合模型的AUC达到了0.947。与单独使用超声或CT进行诊断相比,超声和胸部CT的联合使用展现出了更优的诊断性能。结论采用CNN进行的超声加CT图像深度学习分析具有很高的诊断性能,能够有效地区分乳腺超声图像上的良性和恶性肿瘤,其诊断效能优于单一的影像学方法。
Objective To implement deep learning through convolutional neural networks(CNNs)to distinguish benign and malignant breast masses on US and chest CT images.Methods We retrospectively collected 281 images of benign masses and 155 images of malignant masses from patients who underwent surgical resection and mammotome biopsy to obtain their pathological results.The ResNet-50 architecture was used,and the convolutional block attention module was included to establish deep learning models.The benign and malignant masses were analyzed using US,CT,and US+CT images.The sensitivity,specificity,accuracy,and AUC were calculated.Moreover,the deep learning results were compared with those of manual evaluations.Results The AUC of the ROC curve from ultrasound doctor evaluations on US images was 0.808 and that on CT images was 0.749.In addition,the AUC of the ROC curve from the single US deep learning model was 0.883,while that of the single CT model was 0.871,and that of the combined US+CT model was 0.947.Compared with singlemodality diagnosis and manual evaluations,US combined with chest CT shows the highest diagnostic performance.Conclusion Deep learning analysis of US+CT images using a CNN shows high diagnostic performance and effectively distinguishes benign and malignant breast masses.Its diagnostic accuracy notably outperforms the use of a single imaging modality.
作者
刘晓玲
肖为瀚
覃夏川
LIU Xiao-ling;XIAO Wei-han;QIN Xia-chuan(Department of Ultrasound,Beijing Anzhen Nanchong Hospital of Capital Medical University&Nanchong Central Hospital,Nanchong 637000,Sichuan Province,China;Department of Ultrasound,Chengdu Second People's Hospital,Chengdu 610000,Sichuan Province,China;Department of Ultrasound,the First Affiliated Hospital of Anhui Medical University,Hefei 230022,Anhui Province,China)
出处
《中国CT和MRI杂志》
2025年第7期96-99,共4页
Chinese Journal of CT and MRI
关键词
深度学习
回顾性研究
乳腺肿瘤
CT
超声
Deep Learning
Retrospective Studies
Breast Neoplasms
CT
Ultrasound
作者简介
第一作者:刘晓玲,女,主治医师,主要研究方向:超声造影、介入超声及人工智能超声的应用。E-mail:lxlzlo@126.com;肖为瀚,男,住院医师,主要研究方向:人工智能超声的应用。E-mail:515146462@qq.com;通讯作者:覃夏川,男,主任医师,主要研究方向:超声造影、介入超声及人工智能超声的应用。E-mail:11326636@qq.com。
