超声造影VueBox软件定量分析联合人工智能大语言模型在鉴别C-TIRADS 4类甲状腺小结节良恶性中的价值

Value of quantitative analysis of VueBox software of contrast-enhanced ultrasound combined with artificial intelligence large language model in differential diagnosis of benign and malignant C-TIRADS 4 micro-thyroid nodules

目的:本研究拟探寻超声造影(Contrast-enhanced ultrasound,CEUS)VueBox软件定量分析联合人工智能大语言模型在鉴别C-TIRADS 4类甲状腺小结节良恶性中的价值。方法:回顾性分析2021年3月—2022年11月于我院行常规超声检查发现最大直径≤10 mm的C-TIRADS 4类结节并进一步行CEUS检查,后经甲状腺细针穿刺或手术病理证实的100个患者(100个结节)的资料。根据病理结果分为恶性组(70个结节)和良性组(30个结节),记录两组结节常规超声声像图特征并比较;使用VueBox软件对结节CEUS动态图像进行分析获得时间-强度曲线,记录结节及周围等大等深正常组织共24个VueBox定量参数及差值(△)并比较。采用ChatGPT 3.5大语言模型分析初级医师及高级医师记录的C-TIRADS特征并记录输出结果。评价两组间VueBox定量参数及大语言模型结果,并绘制受试者工作特征(Receptor operation curve,ROC)曲线评估诊断效能。结果:恶性组结节相较于良性组结节,纵横比大于1及边界模糊的比例更高;恶性组结节峰值强度(PE)((11292.37±8608.01)a.u)、结节处PE与正常组织PE差值(△PE)(-13219.32±20922.01)、流入相曲线下面积(WiAUC)((47702.1±44265.22)a.u)、流入相比率(WiR)((4644.31±3857.83)a.u)、流入相灌注指数(WiPI)((10112.57±8468.14)a.u)均低于良性组结节。定量参数PE以10967.66 a.u为截断值诊断效能最高,灵敏度为0.6(0.476~0.713),特异度为0.87(0.684~0.956),准确度为0.68,AUC为0.789(0.695~0.882);大语言模型诊断灵敏度为0.83(0.716~0.905),特异度为0.63(0.439~0.795),准确度为0.77,AUC为0.731(0.616~0.846);CEUS的VueBox定量分析联合大语言模型诊断灵敏度为1(0.935~1),特异度为0.63(0.439~0.795),准确度为0.89,AUC为0.817(0.707~0.926)(P均<0.05)。结论:CEUS VueBox定量参数PE可有效鉴别C-TIRADS 4类甲状腺小结节的良恶性,联合大语言模型可保留诊断特异度的同时提高诊断灵敏度、准确度及AUC。

Objective:To explore the value of quantitative analysis of contrast-enhanced ultrasound(CEUS)VueBox software combined with artificial intelligence large language model in differential diagnosis of benign and malignant C-TIRADS 4 micro-thyroid nodules.Methods:A total of 100 C-TIRADS category 4 micro-thyroid nodules(with a maximum diameter of≤10 mm)of 100 patients detected by conventional ultrasound and further performed CEUS and fine needle aspiration(FNA)or surgery in our hospital from March 2021 to November 2022 were collected retrospectively.According to the pathological results,they were divided into malignant group(70 nodules)and benign group(30 nodules).The conventional ultrasound sonographic features of two groups were recorded and compared.Time-intensity curves were obtained by analyzing the dynamic videos of CEUS using VueBox software,and a total of 24 perfusion parameters,including quantitative parameters of the nodules,the surrounding normal thyroid tissue with equal size and equal depth,and their differences were recorded and compared in two groups.ChatGPT 3.5 large language model was used to analyse the C-TIRADS features recorded by junior and senior radiologist and the output results were recorded.The receptor operation curve(ROC)curve was used to assess the diagnostic efficacy.Results:The proportion of nodules with aspect ratios greater than 1 and blurred borders was higher in the malignant group than that in the benign group.The peak enhancement(PE)(11292.37±8608.01)a.u,the differences in PE between the nodule and normal tissue(ΔPE)(-13219.32±20922.01),wash-in area under the curve(WiAUC)(47702.1±44265.22)a.u,wash-in rate(WiR)(4644.31±3857.83)a.u,wash-in perfusion index(WiPI)(10112.57±8468.14)a.u in the malignant group were lower than those in the be nign group.PE had the highest diagnostic efficacy with a cut-off value of 10967.66 a.u,with a sensitivity of 0.6(0.476~0.713),a specificity of 0.87(0.684~0.956),an accuracy of 0.68 and an AUC of 0.789(0.695~0.882).The diagnostic sensitivity of large language model was 0.83(0.716~0.905),the specificity was 0.63(0.439~0.795),the accuracy was 0.77,and the AUC was 0.731(0.616~0.846),respectively.The diagnostic sensitivity of VueBox analysis combined with the large language model was 1(0.935~1),the specificity was 0.63(0.439~0.795),the accuracy of 0.89,and the AUC of 0.817(0.707~0.926),respectively(all P<0.05).Conclusion:VueBox quantitative parameter PE of CEUS can effectively identify benign and malignant micro-thyroid nodules of C-TIRADS category 4,and the combination with the large language model can improve the diagnostic sensitivity,accuracy,and AUC while preserving the diagnostic specificity.