用于BI-RADS 4类肿块动态超声诊断的人工智能新模型

A novel artificial intelligence model for Breast Imaging Reporting and Data System 4 category breast masses in dynamic ultrasound diagnosis

目的探究一种融合了SAM-YOLOV 5深度学习网络和图像处理技术的人工智能(AI)新模型在乳腺影像报告与数据系统(BI-RADS)4类肿块超声动态视频良恶性分类中的应用。方法回顾性收集2019年5月至2023年6月汕头大学医学院第一附属医院经病理证实的BI-RADS 4类的乳腺肿块患者458例(530个肿块),按7∶3的比例进行模型的训练和测试,分析模型的ROC曲线下面积(AUC)、敏感性、特异性、阳性预测值、阴性预测值。先与单张静态图像下的测试效果进行比较,再与3个传统的深度学习网络以及高、低年资医师组的测试效果进行比较。分析新模型在BI-RADS 4a、4b、4c类肿块中的诊断效能。结果二维超声动态视频在新模型中测试所得到的AUC、敏感性、特异性、阳性预测值、阴性预测值高于使用单张超声静态图像(均P<0.05)。基于二维超声动态视频下,新模型的AUC、敏感性、特异性、阳性预测值、阴性预测值高于3个深度学习网络模型(YOLOV 5、VGG 16、Resnet 50)和低年资医师组(均P<0.05),低于高年资医师组(其中仅特异性、阴性预测值P<0.05)。新模型对BI-RADS 4b类肿块诊断效能最低。结论基于SAM-YOLOV 5深度学习网络和图像处理技术开发的用于BI-RADS 4类乳腺肿块动态超声分类诊断的新模型有较高的诊断价值,有望用于辅助临床诊断。

Objective To investigate the diagnostic performance of a new artificial intelligence(AI)model incorporating SAM-YOLOV 5 deep learning network and image processing techniques for Breast Imaging Reporting and Data System(BI-RADS)4 category breast masses in dynamic ultrasound classification.Methods A total of 530 pathologically proven breast lesions of BI-RADS category 4 in 458 patients were retrospectively collected from May 2019 to June 2023 at the First Affiliated Hospital of Shantou University Medical College.The model was trained and tested at ratio of 7∶3,the area under the ROC curve(AUC),sensitivity,specificity,positive predictive value and negative predictive value of the model were determined.Firstly,the test results of the model were compared with a single static image,then,compared with the three conventional deep learning networks as well as senior and junior radiologists.The diagnostic efficiency of the new model in BI-RADS categories 4a,4b,and 4c masses were analyzed.Results The AUC,sensitivity,specificity,positive predictive value and negative predictive value of the new model based on dynamic ultrasound video were higher than those using a single ultrasound static imaging(all P<0.05).Based on dynamic ultrasound video,the AUC,sensitivity,specificity,positive predictive value and negative predictive value of the new model were significantly higher than those of YOLOV 5,VGG 16,Resnet 50 and the junior group(all P<0.05),lower than the senior group(just specificity and negative predictive value,P<0.05).The diagnostic efficiency of new model for BI-RADS category 4b masses was the lowest.Conclusions Based on the SAM-YOLOV 5 deep learning network and image processing techniques,the new model has a high diagnostic value for breast mass dynamic ultrasound classification and is expected to be used in assisting clinical diagnosis.