| 摘要: |
| 目的：探讨基于关键解剖结构检测的人工智能(AI)在甲状腺超声标准切面(TUSP)识别中的价值。方法：以成人TUSP图像为研究对象，含标准集8978张和实验集1916张。标准集分为训练集和验证集，分别用于训练和验证AI模型识别分类TUSP；实验集以超声专家分类为标准，比较分析AI及不同年资医师组对实验集TUSP、非标准切面(non-standard plane，N-SP)图像识别分类能力的差异。结果:AI对8个TUSP切面的分类准确率从94.7%~99.9%不等，对N-SP的分类准确率为93.8%;AI对TUSP及N-SP各切面识别能力均优于初级医生(P<0.05)；AI对左甲状腺纵切面(longitudinal plane of the left lobe of thyroid, LPLT）、N-SP切面的识别能力与中级医师无显著差异（P=0.468 、P=0.816）,对TUSP其余切面识别能力均优于中级医师（P<0.05)。AI效率显著优于各级医师人工分类（P<0.05)。结论:基于关键解剖结构检测的AI对TUSP、N-SP识别分类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专科培训的辅助方法。 |
| 关键词: 人工智能 超声检查；甲状腺超声标准切面 质量控制 |
| DOI： |
| 投稿时间：2022-12-01修订日期：2023-01-27 |
| 基金项目: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) |
|
| Thyroid ultrasonic standard section recognition by artificial intelligence based on key anatomical structure detection |
| Liu Shunlan,guo minghui,yu zhenggang,lv guorong,liu peizhong,su qichen,he shaozheng |
| () |
| Abstract: |
| Objective: To explore the value of artificial intelligence (AI) based on key anatomical structure in thyroid ultrasound standard plane (TUSP) recognition. Methods: Adult thyroid ultrasound imageswere selected as the research objects, including 8978 TUSP images in standard set and 1916 images in experimental set. The images in standard set were further divided into training set and verifying set for training and verifying the ability of AI in recognizing and classifying TUSP. The images in experimental set including TUSP images and non-standard plane (N-SP) images. Taking the classification results of ultrasound experts as the standard, the differences among AI, junior doctors and intermediate doctors in the recognition and classification on experimental set were compared and analyzed.Results: On experimental set, The classification accuracy of AI for eight TUSP sections ranged were from 94.7% to 99.9%, and the classification accuracy of AI for N-SP was 93.8%.AI was superior to junior doctors in the recognition efficiency of TUSP and N-SP (P<0.05).The ability of AI to recognize longitudinal plane of the left lobe of thyroid (LPLT) and N-SP sections was not significantly different from that of intermediate doctors (P=0.468, P=0.816). And AI performed slightly better than intermediate doctors on other sections of TUSP (P<0.05).The classification efficiency of AI was significantly better than the that of artificial classification (P<0.05).Conclusion :AI based on key anatomical structure detection has high accuracy and efficiency for TUSP and N-SP recognition and classification , which can be used as an auxiliary method for thyroid ultrasound image quality control and specialized training. |
| Key words: artificial intelligence ultrasonography thyroid ultrasound standard plane quality control |
