探究不同乳腺实质区域的全视野数字化乳腺摄影影像组学对BI-RADS 4类病变鉴别的价值

Exploring the Value of FFDM Radiomics in Different Parenchymal Regions of the Breast for the Differentiation of BI-RADS 4 Lesions

目的 探究基于全视野数字化乳腺摄影(FFDM)的瘤内、瘤周以及对侧实质区域的影像组学模型鉴别乳腺BI-RADS 4类病灶良恶性的价值。方法 回顾性分析2020年11月至2021年8月行FFDM检查被诊断为BI-RADS 4类的117例患者。分别勾画病灶的瘤内ROI、瘤周ROI和对侧乳腺实质ROI,提取并筛选影像组学特征,运用支持向量机建立3个不同区域的影像组学模型,选取性能较好的模型用于构建多区域影像组学联合模型。同时对常见的临床数据进行筛选并建立临床模型,最后构建临床-影像组学联合模型。模型的性能由AUC和准确率等指标评估。结果 共51例良性、66例恶性病例,在3个不同区域构建了单区域影像组学模型,除了对侧实质模型,其他均可以有效鉴别BI-RADS 4类病灶的性质,其中联合模型优于单区域影像组学模型和临床模型,并且优于放射科医师。瘤内模型、瘤周模型、对侧实质模型和临床-影像组学联合模型在测试集的AUC分别为0.80、0.74、0.52、0.87。结论 基于FFDM的瘤内、瘤周影像组学模型可以鉴别乳腺BI-RADS 4类病变的性质,由其构成的联合模型可以进一步提高对BI-RADS 4类病变的诊断效能,但对侧实质纹理模型无法鉴别乳腺病变的良恶性。

Objective To investigate the value of FFDM-based intratumoral,peri-tumoral and contralateral parenchymal radiomics models in identifying benign and malignant BI-RADS 4 lesions in the breast.Methods Patients diagnosed with BI-RADS 4 on FFDM from November 2020 to August 2021 were retrospectively analyzed.The intratumoral ROI,peri-tumoral ROI and contralateral breast parenchyma ROI of the lesions were outlined separately,and the radiomics features were extracted and screened,and the SVM was used to build three different regional radiomics models,and the model with better performance was selected for the construction of a combined multi-regional radiomics model.At the same time,common clinical data were screened and clinical models were built,and finally the joint clinical-radiomics model was constructed.The performance of the models was evaluated by indicators such as AUC and accuracy rate.Results A total of 51 benign cases and 66 malignant cases with single-region radiomics models constructed in 3 different regions could effectively identify the character of BI-RADS 4 lesions,except for the contralateral parenchymal model,in which the combined model was superior to the single-region radiomics model and clinical model,and was superior to the radiologist.The AUCs of the intratumoral,peri-tumoral,contralateral parenchymal,and combined clinical-radiomics models in the test set were 0.80,0.74,0.52,and 0.87,respectively.Conclusion The FFDM-based intratumoral and peri-tumoral radiomics models can identify the character of BI-RADS 4 lesions,and the combined model composed of them can further improve the diagnostic efficiency of BI-RADS 4 lesions,but the contralateral parenchymal model cannot identify the benignity and malignancy of breast lesions.