基于深度学习模型的S-Detect技术在甲状腺良恶性结节鉴别诊断中的应用价值

Application value of S-Detect technology based on deep learning model in differential diagnosis of benign and malignant thyroid nodules

目的:探讨基于深度学习模型的超声人工智能辅助诊断技术(S-Detect技术)在甲状腺结节诊断中的临床应用价值。方法:选取2019年10月至2020年5月在河南省肿瘤医院治疗并行超声检查的甲状腺结节患者(共183例患者,183个病灶),分别由超声医生及S-Detect技术对其超声声像图特征进行分析,比较二者对不同超声特征诊断的一致性,并以病理结果为金标准,评估S-Detect技术在甲状腺结节鉴别诊断中的临床应用价值。结果:183例甲状腺结节经病理证实良性68例,恶性115例,S-Detect技术对本组数据诊断的敏感性、特异性、准确率分别为84.35%、86.76%、85.25%;高年资医师组诊断的敏感性、特异性、准确率分别为90.43%、92.65%、91.26%,低年资医师组诊断的敏感性、特异性、准确率分别为76.52%、82.35%、78.69%,在可以评估的5个超声声像图特征中,结节成分、回声及纵横比3个指标超声医师与S-Detect诊断的一致性较高,结节形状及边缘2个指标超声医师与S-Detect诊断的一致性较差。结论:基于深度学习模型的S-Detect技术诊断甲状腺结节的准确率较高,有助于提高低年资医师诊断的准确性。

Aim:To explore the clinical application value of S-Detect,an ultrasound artificial intelligence-aided diagnosis technology based on deep learning model,in thyroid nodule diagnosis.Methods:A total of 183 patients with thyroid nodules underwent ultrasound examination in Henan Cancer Hospital from October 2019 to May 2020 were selected.The consistency of ultrasonographic features respectively analyzed by ultrasound doctors and S-Detect technology was compared.Taking the pathological results as the gold standard,this clinical value of S-Detect in differential diagnosis of thyroid nodule was evaluated.Results:Sixty-eight cases of thyroid nodules were pathologically proved benign,and 115 cases were malignant.The sensitivity,specificity and accuracy of S-Detect in the diagnosis of thyroid nodules were 84.35%,86.76%and 85.25%,respectively.The sensitivity,specificity and accuracy of diagnosis in the senior physician group were 90.43%,92.65%and 91.26%,while those in the junior physician group were 76.52%,82.35%and 78.69%.Among the five ultrasound image features that could be evaluated,the consistency between ultrasound physicians and S-Detect diagnosis was higher with respect to mass composition,echo and aspect ratio.The consistency between ultrasound physicians and S-Detect diagnosis was poorer in the diagnosis of tumor shape and margin.Conclusion:The diagnostic accuracy of S-Detect based on deep learning model is high,which is helpful to improve the diagnostic accuracy of thyroid nodules for junior doctors.