基于U-Det模型对肺内小结节CT图像良恶性鉴别的价值分析

Analysis of the Value of U-Det Model in Differentiating Benign from Malignant Small Pulmonary Nodules on CT Images

目的探讨U-Det模型对肺内小结节CT图像良恶性鉴别的价值。方法选取2021年6月至2022年1月于我院经病理检查的150例肺内小结节患者样本图片,留取恶性样本(n=104)及非恶性样本(n=46),扩增至各800张。按照7∶3比例随机分为训练集样本(n=1120)和验证集样本(n=480)。根据训练样本对预训练的卷积神经网络架构ResNet50进行训练,建立卷积神经网络计算机辅助系统,测试筛选肺内小结节恶性病变方面的能力;同时选取LUNA16的1400张病理图片作为测试集,测试U-Det模型的诊断效能。结果U-Det模型中训练样本的平均损失率为0.126%±0.046%,验证样本的平均损失率为0.135%±0.053%。U-Det模型中训练样本的平均准确度为88.42%±4.21%,验证样本的平均准确度为89.01%±4.09%。受试者工作特征曲线显示,U-Det、U-Net和ResNet50模型预测准确度递减(P<0.05);LUNA16测试集下U-Det模型的诊断准确度、敏感度、特异性、阳性预测值、阴性预测值最高,U-Net次之,ResNet50最低。结论U-Det模型对肺内小结节CT图像良恶性鉴别价值较高,可将其用于肺内小结节良恶性诊断。

Objective To investigate the value of U-Det model in differentiating benign from malignant pulmonary nodules on CT images.Methods Samples of 150 patients with small pulmonary nodules who underwent pathological examination in our hospital from June 2021 to January 2022 were selected.Malignant samples(n=104)and non-malignant samples(n=46)were retained and increased to 800 samples each.According to the ratio of 7∶3,the samples of training set(n=1120)and verification set(n=480)were randomly divided into two groups.According to the training samples,the pre-trained convolutional neural network architecture ResNet50 was trained,and the convolutional neural network computer aided system was established to test the ability of screening malignant lesions of small pulmonary nodules.Meanwhile,1400 pathological images of LUNA16 were selected as the test set to test the diagnostic value of U-Det model.Results The average loss rate of training samples in U-Det model was 0.126%±0.046%,and the average loss rate of verification samples was 0.135%±0.053%.The average accuracy of training samples and verification samples in U-Det model was 88.42%±4.21%and 89.01%±4.09%respectively.The receiver operating characteristic curve showed that the prediction accuracy of U-Det,U-Net and ResNet50 models decreased successively(P<0.05).The diagnostic accuracy,sensitivity,specificity,positive and negative predictive values of U-Det model in LUNA16 test set were the highest,followed by U-Net and ResNet50.Conclusion U-Det model has a high value in the differential diagnosis of benign from malignant pulmonary nodules on CT images.It can be used in the diagnosis of benign from malignant small pulmonary nodules.