基于CT深度学习鉴别胸腺瘤组织学分型的临床应用价值

Clinical application value of CT deep learning in identifying histological types of thymoma

目的研究基于CT深度学习鉴别胸腺瘤组织学分型的临床应用价值。方法回顾性分析2014年1月—2023年3月于本院经术后病理证实的179例胸腺瘤患者的临床资料,按7︰3比列将研究对象分为训练集(n=125)和验证集(n=54)。本研究采用ResNet50作为卷积神经网络模型,提取深度学习特征,并运用主成分分析(PCA)、相关性分析和最小绝对收缩和选择算子(LASSO)算法筛选出最优深度学习特征,建立鉴别胸腺瘤组织学分型的深度学习模型(DTL Signature)。利用Logistic回归构建临床模型、深度学习模型和联合模型。绘制受试者工作特征曲线(ROC),并计算曲线下面积(AUC),评估3种模型鉴别胸腺瘤组织学分型的临床效能。结果筛选出11个CT平扫深度学习特征,并建立深度学习模型,在训练集和验证集中,其受试者工作曲线(ROC)下面积(AUC)分别为0.877、0.795,均高于临床模型的AUC,但差异均无统计学意义(Z=1.903、1.033,P均>0.05)。在训练集和验证集中,临床模型与深度学习模型构建联合模型,其ROC的AUC分别为0.890、0.841,均高于临床模型,且差异有统计学意义(Z=2.647、3.041,P均<0.05)。结论基于CT平扫的深度学习联合模型有利于鉴别胸腺瘤组织学分型,临床获益较高,值得临床推广应用。

Objective To study the clinical application value of CT deep learning in identification of histological types of thymoma.Methods The clinical data of 179 patients suffering from thymoma confirmed by postoperative pathology in the first affiliated hospital of Wannan Medical College from January 2014 to March 2023 were retrospectively analyzed.The subjects were divided into training set(n=125)and verification set(n=54)according to 7:3 ratio.Resnet50 was selected as the convolutional neural network model to extract deep learning features.Principal component analysis,correlation analysis and minimum absolute contraction and least absolute shrinkage and selection operator(LASSO)were used to screen out the optimal deep learning features,and a deep learning model(DTL Signature)was constructed for the identification histological types of thymoma.Logistic regression was used to construct clinical model,deep learning model and association model respectively.Receiver operating characteristic(ROC)curve was plotted and area under the curve(AUC)was calculated to evaluate the clinical efficacy of the three models in differentiating histological types of thymoma.Results Eleven characteristics of CT scan deep learning were screened out,and deep learning model was set up.In the training set and validation set,the area under the receiver-operating curve(AUC)were 0.877 and 0.795 respectively,both were higher than that of clinical models,however,there were no statistically significant difference(Z=1.903、1.033,P>0.05).In the training set and validation set,a combination model was constructed by clinical model and deep learning model,the AUC were 0.890,0.841,both were higher than that of clinical model,and the differences were statistically significant(Z=2.647、3.041,P<0.05).Conclusions The combined deep learning model based on CT plain scan is beneficial to the identification of histological types of thymoma,with high clinical benefits,and worthy of clinical application.