基于动态增强磁共振成像的影像组学和不同CNN的深度学习对乳腺良恶性病变的诊断价值

Diagnosis of benign and malignant breast lesions based on DCE-MRI by using radiomics and deep learning with different networks

目的:探讨基于动态增强磁共振成像(DCE-MRI)的影像组学和深度学习对乳腺良恶性病变的诊断价值。方法:选取2017年1月至2018年6月经手术或穿刺病理证实的乳腺恶性病变患者93例、良性59例。所有患者均术前行乳腺MRI平扫及增强检查。通过计算获得早期流入信号增强比、最大增强信号比和流出斜率相对应的参数图。采用基于纹理和强度直方图的影像组学分析和5个卷积神经元网络(CNN)模型(ResNet50、VGG16、VGG19、Xception和InceptionV3)的深度学习进行诊断分析,并采用包含不同量肿瘤周围组织的五种不同输入边界框分析。结果:影像组学的诊断准确度均值为80%。不同输入边界框的ResNet50模型中,含少量肿瘤周围组织的最小边界框诊断效能最高,高于1.5倍和2.0倍边界框分析(均P<0.01);ResNet50(93%)、Xcep-tion(94%)和InceptionV3(93%)的准确度高于VGG16(80%)和VGG19(79%)(均P<0.01)。结论:深度学习对乳腺良恶性病灶的诊断效能优于影像组学,含少量肿瘤周围组织分析的诊断效能高于仅包含肿瘤组织和包含较多肿瘤周围组织。

Objective:To evaluate and compare the performance of radiomics and deep learning in the diagnosis of benign and malignant breast lesions based on DCE-MRI.Methods:A total of 152 patients receiving breast MRI for diagnosis were analyzed,including 93 patients with malignant cancers,and 59 patients with benign lesions.Three DCE parametric maps corresponding to early wash-in signal enhancement(SE),maximum signal enhancement,and wash-out slope were generated.Radiomics analysis based on texture and intensity histogram,and deep learning using 5 networks(ResNet50,VGG16,VGG19,Xception,InceptionV3),were performed for differential diagnosis.Results:The accuracy of radiomics was 80%;the smallest bounding box obtained higher diagnostic accuracy than that with tumor only and 1.2 times box which was not significant(both P>0.05),and also higher than 1.5 times box and 2.0 times box(both P<0.01);the accuracy of ResNet50(93%),Xception(94%),InceptionV3(93%)was significantly higher than VGG16(80%),VGG19(79%)(both P<0.01).Conclusion:CNN achieved better diagnostic performance in the diagnosis of benign and malignant breast lesions.The smaller bounding box containing the tumor with small amount of per-tumor tissue had the higher diagnostic accuracy than that with the tumor only and than larger bounding box.