摘要
目的探讨基于关键解剖结构检测的人工智能(AI)模型在甲状腺超声标准切面(TUSP)识别中的应用价值。方法以成人TUSP和非标准切面(N-SP)图像为研究对象,含标准集8978张和实验集1916张;其中标准集分为训练集8178张和验证集800张,分别用于训练和验证AI模型识别TUSP;以超声专家团队识别为标准,比较初级医师、中级医师及AI模型识别实验集TUSP和非标准切面(N-SP)的诊断效能;同时收集AI模型及不同年资医师识别实验集图像累计耗时及平均每张图像耗时,并对其进行比较。结果AI模型识别实验集8个TUSP即甲状腺峡部横切面(TPTI)、甲状腺峡部纵切面(LPTI)、左甲状腺上极横切面(UTPLT)、左甲状腺下极横切面(DTPLT)、右甲状腺上极横切面(UTPRT)、右甲状腺下极横切面(DTPRT)、左甲状腺纵切面(LPLT)及右甲状腺纵切面(LPRT)的准确率为94.7%~99.9%,识别N-SP的准确率为93.8%;AI模型识别8个TUSP和N-SP的曲线下面积(AUC)均大于初级医师,差异均有统计学意义(均P<0.05);AI模型识别LPTI、UTPLT、DTPLT、UTPRT、DTPRT、LPRT的AUC均大于中级医师,差异均有统计学意义(均P<0.05);中级医师识别TPTI、UTPRT、DTPRT、LPLT、LPRT、N-SP的AUC均大于初级医师,差异均有统计学意义(均P<0.05);其余两两比较差异均无统计学意义。AI模型识别实验集TUSP图像累计耗时及平均每张图像耗时均少于不同年资医师人工识别,专家团队识别累计耗时及平均每张图像耗时均少于中级、初级医师,中级医师识别累计耗时及平均每张图像耗时均少于初级医师,差异均有统计学意义(均P<0.05)。结论基于关键解剖结构检测的AI模型识别TUSP具有较高准确性和效率,可作为甲状腺超声图像质量控制和规范化培训的辅助工具。
Objective To explore the clinical value of artificial intelligence(AI)model based on key anatomical structure detection in thyroid ultrasound standard plane(TUSP)recognition.Methods Adult TUSP and non-standard section(N-SP)ultrasound images were selected as the research objects,including 8978 images in standard set and 1916 images in experimental set.The standard set was further divided into a training set of 8178 images and a validation set of 800 images,for training and verifying the ability of AI model in recognizing and classifying TUSP images.Taking the classification of ultrasound experts as the standard,the diagnostic efficacy of junior,intermediate physicians,and AI model in identifying TUSP and non-standard plane(N-SP)images in experimental set were compared.The cumulative time consumption and average time consumption per image of AI model and different seniority physicians in identifying experimental set of images were compared.Results In experimental set,the classification accuracy of AI model for transverse plane of thyroid isthmus(TPTI),longitudinal plane of thyroid isthmus(LPTI),upside of the transverse plane of the left lobe of thyroid(UTPLT),downside of the transverse plane of the left lobe of thyroid(DTPLT),upside of the transverse plane of the right lobe of thyroid(UTPRT),downside of the transverse plane of the right lobe of thyroid(DTPRT),longitudinal plane of the left lobe of thyroid(LPLT),longitudinal plane of the right lobe of thyroid(LPRT)ranged from 94.7%to 99.9%,and the classification accuracy of AI model for N-SP was 93.8%.The AUC of the 8 TUSP and N-SP sections identified by AI model were higher than those of the primary physician,and the differences were statistically significant(all P<0.05).The AUC of LPTI,UTPLT,DTPLT,UTPRT,DTPRT,and LPRT identified by AI model were higher than those of intermediate physicians,and the differences were statistically significant(all P<0.05).The AUC of TPTI,UTPRT,DTPRT,LPLT,LPRT,and N-SP identified by intermediate physicians were higher than those of junior physicians,and the differences were statistically significant(all P<0.05).There were no statistically significant difference between the other two comparisons.The cumulative and average time spent on image classification by AI model in the TUSP experimental set were significantly less than those of manual recognition by different seniorty physicians.The cumulative and average time spent on experts recognition were less than those of intermediate and junior physicians,and the cumulative and average time spent on image recognition by intermediate physicians were less than those of junior physicians,with statistically significant differences(all P<0.05).Conclusion The AI model based on key anatomical structure detection has high accuracy and efficiency for TUSP recognition,which can be used as an auxiliary method for thyroid ultrasound image quality control and specialized training.
作者
柳舜兰
郭明辉
喻正纲
柳培忠
苏淇琛
何韶铮
吕国荣
LIU Shunlan;GUO Minghui;YU Zhenggang;LIU Peizhong;SU Qichen;HE Shaozheng;LV Guorong(Department of Ultrasound,the Second Affiliated Hospital of Fujian Medical University,Fujian 362000,China)
出处
《临床超声医学杂志》
CSCD
2023年第5期372-377,共6页
Journal of Clinical Ultrasound in Medicine
基金
教育部泉州医学高等专科学校母婴健康服务应用技术协同创新中心经费资助项目[闽科教(2017)49号]
泉州市科技计划资助项目(2019C076R)。
关键词
超声检查
人工智能
甲状腺超声标准切面
质量控制
Ultrasonography
Artificial intelligence
Thyroid ultrasound standard plane
Quality control
