摘要
目的 开发一种基于深度学习的上腹部超声标准切面识别与分类算法,以提高超声检查的标准化程度和诊断准确度,减少医师经验依赖性,优化超声检查流程,提高医疗资源利用效率。方法 选择2024年5月至2025年1月在北京清华长庚医院就诊行肝胆胰脾超声检查的患者183例,其中男性100例,女性83例;年龄22~56岁,平均年龄39.47岁;身体质量指数18.90~25.40 kg/m^(2),平均身体质量指数21.79 kg/m^(2)。行超声视频样本。采用改进的SE-net网络构建标准切面分类模型,并使用YOLOv8进行脏器自动分割。图像预处理包括归一化、固定区域裁剪,并采用旋转、缩放、对比度调整等数据增强策略,以提高模型的泛化能力。训练过程中使用交叉熵损失函数,并在预测阶段结合测试时增强(TTA)策略优化结果。模型性能通过准确度、召回率、特异度、精确率、Dice系数等指标进行评估。结果 所有标准切面图像共计3 492帧。经腹主动脉纵断面388帧;经下腔静脉纵断面84帧;经门静脉右前叶支和胆囊颈部右肋间斜断面110帧;经第二肝门肋缘下斜断面100帧;肝-肾纵断面447帧;门静脉左支矢状部断面225帧;胆囊长轴断面956帧;胆囊短轴断面100帧;肝门部肝外胆管断面32帧;胰腺长轴断面412帧;胰头短轴断面11帧;脾脏肋间前斜冠状断面627帧。该模型成功实现了对12个上腹部超声标准切面的自动识别与分类。模型在标准切面识别任务中的总体性能指标为:准确度85.07%,召回率72.61%,特异度98.59%,精确率81.05%。表明该算法能够有效识别上腹部超声标准切面,提高诊断可靠性。在脏器分割任务中,YOLOv8在肝脏、胆囊、右肾的分割性能较高(Dice=0.87、0.90、0.87),但在胰腺的识别上表现较差(Dice=0.38)。结论 实验提出的深度学习模型能够高效、准确地识别上腹部超声标准切面,并对脏器进行自动分割,提升超声检查的标准化程度和诊断一致性,降低对医师经验的依赖,提高超声检查的工作效率。未来研究可进一步优化模型结构,提高对胰腺等难识别脏器的分割能力,并结合多模态影像进行深度优化,以提升模型在实际临床应用中的可靠性和适用性。
Objective To develop an upper abdominal ultrasound standard section recognition and classification algorithm based on deep learning,and enhance the standardization and diagnostic accuracy of ultrasound examinations,reduce reliance on operator experience,optimize the ultrasound examination process,and improve the efficiency of medical resource utilization.Methods From May 2024 to January 2025,a total of 183 patients who underwent hepatobiliary,pancreatic,and splenic ultrasound examinations were enrolled,which included 100 males and 83 females,aged 22-56 years old with mean age of 39.47 years old;body mass index was 18.90-25.40 kg/m^(2) with mean of 21.79 kg/m^(2).The ultrasound video samples were collected.The modified squeeze-and-excitation networks(SE-net)were used to construct the standard plane classification model,and the organs were automatically segmented by YOLOv8.The image preprocessing data of normalization,fixed-region cropping,rotation,scaling,and contrast adjustment and other data enhancement strategies were adopted to improve the generalization ability of the model.The cross-entropy loss function was used in the training process,and the test-time augmentation(TTA)strategy optimization results were combined in prediction stage.The indicators such as model performance accuracy,recall,specificity,accuracy,and Dice coefficient were evaluated.Results A total of 3492 standard section images were obtained.A total of 388 images were transabdominal aortic longitudinal sections,84 longitudinal sections via inferior vena cava,110 oblique sections via right anterior lobe branch of portal vein and gallbladder neck right rib,100 subcostal oblique sections of the second hepatic hilum,447 hepatorenal longitudinal sections,225 sagittal sections of the left branch of portal vein,956 longitudinal sections of the gallbladder,100 transverse sections of the gallbladder,32 sections of extrahepatic bile duct at hepatic hilum,412 longitudinal sections of the pancreas,11 transverse sections of caput pancreatic,splenic intercostal anterior oblique coronal sections and 627 anterior oblique coronal sections of spleen intercostals.The model successfully achieved automatic recognition and classification of 12 standardized upper abdominal ultrasound planes.The overall performance index of the standard plane recognition were as follows:accuracy 85.07%,recall 72.61%,specificity 98.59%,accuracy 81.05%,which showed that the algorithm could effectively identify standardized upper abdominal ultrasound planes and improve diagnostic reliability.In the organ segmentation task,YOLOv8 demonstrated high performance in segmenting liver,gallbladder,and right kidney(Dice=0.87,0.90,and 0.87,respectively),but exhibited poor performance in pancreatic recognition(Dice=0.38).Conclusion It is demonstrated that the deep learning model efficiently and accurately identifies standardized upper abdominal ultrasound planes and performs automatic organ segmentation,thereby enhancing the standardization and diagnostic consistency of ultrasound examinations,reducing dependence on physician experience,and improving efficiency of ultrasound examination.Future research could further optimize the model structure,improve the segmentation ability of difficult-to-identify organs such as pancreas,and combine multimodal images for deep optimization to improve the reliability and applicability of the model in clinical practice.
作者
张哲元
张华斌
冯月
穆金宝
周璐
郑忠青
王邦茂
王修明
ZHANG Zhe-yuan;ZHANG Hua-bin;FENG Yue;MU Jin-bao;ZHOU Lu;ZHENG Zhong-qing;WANG Bang-mao;WANG Xiu-ming(Department of Ultrasound,Beijing Tsinghua Changgung Hospital,School of Clinical Medicine,Tsinghua University,Beijing 102218,China;Tianjin Yujin Artificial Intelligence Medical Technology Co.,Ltd,Tianjin 300384,China;Department of Gastroenterology,Tianjin Medical University General Hospital,Tianjin 300052,China)
出处
《生物医学工程与临床》
2025年第4期473-479,共7页
Biomedical Engineering and Clinical Medicine
关键词
深度学习
超声检查
标准切面
压缩和激励网络
YOLOv8
deep learning
ultrasound examination
standard plane
squeeze-and-excitation networks(SE-net)
YOLOv8
