基于深度学习的乳腺数字化X线BI-RADS密度分类的研究

摘要

目的 为了确保基于ACR BI-RADS乳腺X线密度评估的一致性和准确性, 本研究构建基于深度学习的乳腺数字化X线密度的自动分类模型, 使之达到对乳腺密度的精准分类.方法 在研究中, 构建并优化基于深度学习的卷积神经网络 (CNN) 的经典模型ResNet 50.收集本院于2015-08～2018-02间行全数字化乳腺摄影图像18 152幅, 由两位有经验的放射科医师根据ACR BI-RADS标准对图像的乳腺密度进行评估.各自经微调的分类模型分别在小数据集 (4 000幅) 和原始数据集 (18 152幅) 对乳腺密度的分类进行评估, 得到相应的分类准确性, 以受试者工作特性曲线和曲线下面积评估模型的分类性能.结果 CNN模型在小数据集训练时, 各类的分类准确性分别为a类91％, b类86％, c类84％, d类90％;当在原始数据集训练时, a类和d类的分类准确性无明显变化, b类和c类的准确性分别为89％和88％, 随着数据量的增加, 准确率明显提高, 比较AUC发现分类性能明显改善.结论 基于深度学习的卷积神经网络 (CNN) 分类模型能以较高的准确率对乳腺密度进行分类, 在临床工作中, 可协助放射科医师对乳腺密度进行准确、一致的分类.%Objective To construct an automatic classification model of mammographic digital X-ray density based on deep learning for accurate classification of mammographic density for ensuring the consistency and accuracy based on ACR BI-RADS mammographic density assessment. Methods The classic model ResNet 50 based on deep learning convolutional neural network (CNN) was constructed and optimized. A total of 18 152 images of all-digital mammography from August 2015 to February 2018 were collected, and the breast density of all the images was assessed by two experienced radiologists according to the ACR BI-RADS criteria. To obtain the classification accuracy, the fine-tuned classification models were used to evaluate the classification of breast density in small datasets (n = 4 000) and original datasets (n = 18 152), respectively. The classification performance of the model was evaluated by the receiver operating characteristic curve and the area under the curve. Results When the CNN model was trained in small data sets, the classification accuracy of each type was 91％ for class a, 86％ for class b, 84％ for class c, and 90％ for class d. When training in the original data sets, the classification accuracy for the class a and d did not change significantly. The accuracy for b and c was 89％ and 88％, respectively. With the increase of data volume, the accuracy was significantly improved. The AUC results also showed that the classification performance was significantly improved. Conclusion The deep learning-based convolutional neural network (CNN) classification model can classify breast density with high accuracy. In clinical work, it can assist radiologists to accurately and consistently classify breast density.