MobileNetV2对乳腺X线BI-RADS 4类病变的降级作用初步研究

A Preliminary Study of MobileNetV2 to Downgrade Classification in Mammographic BI-RADS 4 Lesions

目的 探讨基于MobileNetV2深度迁移学习(DTL)对乳腺X线摄影乳腺影像报告和数据系统(BI-RADS)4类病变降级分类的价值.方法 将良性组、恶性组图像分别按照9:1随机分为训练集(良性组9346幅,恶性组4421幅)和测试集(良性组1038幅,恶性组491幅).通过模型微调构建基于MobileNetV2的DTL模型,并对9346幅良性组和4421幅恶性组乳腺X线图像进行学习,另外搜集由5位影像科医师报告的乳腺X线BI-RADS 4类病变患者共382例作验证集.每个病变均选择头尾位(CC位)和内外斜位(MLO位)两幅图像进行验证,如有1幅图像归类正确,则判断为该例归类正确.以验证集准确率、召回率、F1评分及受试者工作特征曲线(ROC)曲线下面积(AUC)作为DTL模型的性能指标.结果 模型在训练集和测试集准确率分别为100%、98%.在验证集准确率、召回率、F1评分及AUC分别为0.91、0.91、0.91和0.91.模型对BI-RADS 4A、4B、4C类病变降级比例分别为87.7%、80.2%和75.2%.对BI-RADS 4类病变总体降级比例为81.9%,且DTL模型与病理组织学在对乳腺X线摄影良恶性病变的分类诊断结果差异无统计学意义(P=0.206).结论 基于MobileNetV2的DTL模型是乳腺X线摄影BI-RADS 4类病变降级的有效方法.

Objective To explore the potential value of deep transfer learning based on MobileNetV2 to downgrade classification in mammographic BI-RADS 4 lesions.Methods All breast lesions were pathologically confirmed.Dataset of breast mammography images was randomly divided into two subsets,a training set(9346 benign and 4421 malignant)and testing set(1038 benign and 491 malignant),in a proportion of 9∶1.Transfer learning based on MobileNetV2 model was performed for the mammography images.Mammography images from another 382 patients with BI-RADS 4 breast lesions(84 malignant and 298 benign)were employed as validation set.2 images(CC and MLO)from each lesions were tested.Trials were categorized as correct if the judgement(≥1)was corrected.We used precision,recall,F1 scores and the area under the receiver operating characteristic curve(AUC)as performance metrics of validation set.Results Precision,recall,F1 scores and AUC were 0.91,0.91,0.91 and 0.91,respectively.The proportion of downgraded lesions was 87.7%、80.2%and 75.2%for category 4A、4B、4C,respectively.The proportion of downgraded lesions was 81.9%for category 4.The differences between the DTL model and histopathology were not statistically significant(P=0.206).Conclusion DTL model based on MobileNetV2 model can be used as an effective approach to downgrade classification in mammographic BI-RADS 4 lesions.