可解释性人工智能超声影像特征风险模型预测甲状腺乳头状癌颈部淋巴结转移的价值

Value of explainable artificial intelligence ultrasound characteristic risk model in predicting cervical lymph node metastasis of papillary thyroid carcinoma

目的构建甲状腺乳头状癌(PTC)风险特征可解释性人工智能(AI)模型,并探讨其联合临床特征预测PTC患者颈部淋巴结转移(CLNM)的价值。方法回顾性收集西安交通大学第二附属医院2021年1月至2022年9月行甲状腺切除术及颈部淋巴结清扫术后病理证实的PTC患者422例,共422个结节,按7∶3比例随机分为训练集和测试集,通过传统机器学习方法提取与PTC风险特征高度相关的影像组学特征,并建立风险特征概率最优智能预测模型,再联合临床特征构建预测PTC患者CLNM的风险模型,并通过绘制ROC曲线,计算曲线下面积(AUC)评估各模型的诊断效能。结果在测试集PTC风险特征AI可解释模型中,基于逻辑回归分类的钙化智能诊断模型表现出最高的诊断效能,AUC为0.87(P<0.05)。对比于单独PTC超声风险特征概率模型,其联合临床特征的列线图综合模型在预测PTC患者CLNM中表现出更高诊断效能,其AUC为0.97,诊断临界值为0.15,对应的准确性、敏感性及特异性分别为92.65%、92.76%及92.54%(均P<0.05)。结论本研究构建的PTC超声AI模型输出的可解释性风险特征结合临床特征能够有效预测PTC患者的CLNM,进而为医生决策PTC患者治疗方案提供有效信息。

Objective To construct an explainable artificial intelligence(AI)model of risk characteristics of papillary thyroid carcinoma(PTC),and to explore its value of it combined with clinical features in predicting cervical lymph node metastasis(CLNM)in PTC patients.Methods From January 2021 to September 2022,422 patients(422 nodules)with pathologically confirmed PTC underwent thyroidectomy and neck lymph node dissection in the Second Affiliated Hospital of Xi′an Jiaotong University were retrospectively collected,the patients were randomly divided into training set and test set according to the ratio of 7∶3.Ultrasonographic features highly correlated with PTC risk characteristics were extracted by traditional machine learning method,and an intelligent prediction model with optimal probability of risk characteristics was established.Then,a risk model for predicting CLNM of PTC patients was constructed in combination with clinical features.The diagnostic effectiveness of the model was evaluated by drawing a ROC curve and calculating the area under curve(AUC).Results In the AI explaineable model of PTC risk characteristics in the test set,the intelligent diagnosis model of calcification based on logistic regression classification showed the highest diagnostic efficiency,with an AUC of 0.87(P<0.05).Compared with the probability model of risk characteristic of PTC alone,the comprehensive model combined with clinical characteristics showed higher diagnostic efficiency in predicting CLNM of PTC patients,with AUC of 0.97,diagnostic critical value of 0.15,corresponding accuracy,sensitivity and specificity of 92.65%,92.76%and 92.54%,respectively(all P<0.05).Conclusions The explaineble risk characteristics of PTC AI model combined with clinical features can effectively predict the cervical lymph node metastasis of PTC,and then provide effective information for clinical decision-making of PTC patients.