Recently, computer-aided diagnosis (CAD) systems powered by deep learning (DL) algorithms have shown excellent performance in the evaluation of digital mammography for breast cancer diagnosis. However, such systems typically require pixel-level annotations by expert radiologists which is prohibitively time-consuming and expensive. Medical institutes would wonder if a high-performance breast cancer CAD system can be trained by exploring their own huge amount of historical imaging data and corresponding diagnosis reports, without additional annotations workload of their radiologists. In this study, we show that a DL classification model trained on historical mammograms with only image-level pathology labels (which can be automatically extracted from medical reports) can achieve surprisingly good diagnostic performance on newly incoming exams compared with experienced radiologists. A DL model called DenseNet was trained and cross-validated with 5979 historical exams acquired before September 2017 with biopsy-verified pathology and tested with 1194 newly obtained cases after that. For both cross-validation and test sets, the ROCs generated by DL predictions were above the ROCs generated by ratings from radiologists. For the suspicious cases which radiologists suggest biopsy (BI-RADS category 4 and 5), the DL model can reject 60% of false biopsies on benign breasts while keeping 95% sensitivity. For the mammograms based on which radiologists were not able to make a diagnosis (BI-RADS 0), the DL model still achieved an AUC score of 79%. Moreover, the model is able to localize lesions on mammograms although such information was not provided in the training phase. Finally, the impact of input image resolution and different DL model architectures on the diagnostic accuracy were also presented and analyzed.

最近，由深度学习 (DL) 算法提供支持的计算机辅助诊断 (CAD) 系统在评估用于乳腺癌诊断的数字乳房 X 光检查方面表现出色。然而，这样的系统通常需要由放射科专家进行像素级注释，这既耗时又昂贵。医疗机构会想知道是否可以通过探索自己的大量历史影像数据和相应的诊断报告来训练高性能的乳腺癌 CAD 系统，而无需放射科医生的额外注释工作量。在这项研究中，我们表明，与经验丰富的放射科医生相比，仅在具有图像级病理标签（可以从医疗报告中自动提取）的历史乳房 X 光照片上训练的 DL 分类模型可以在新进入的检查中获得令人惊讶的良好诊断性能。一个名为 DenseNet 的 DL 模型经过训练，并与 2017 年 9 月之前获得的 5979 次历史检查进行交叉验证，并使用活检验证的病理学，并在此后对 1194 个新获得的病例进行测试。对于交叉验证和测试集，DL 预测生成的 ROC 高于放射科医生评级生成的 ROC。对于放射科医生建议活检的可疑病例（BI-RADS 类别 4 和 5），DL 模型可以拒绝 60% 的良性乳房假活检，同时保持 95% 的灵敏度。对于放射科医师无法做出诊断的乳房 X 光照片（BI-RADS 0），DL 模型仍然达到了 79% 的 AUC 评分。此外，该模型能够在乳房 X 光照片上定位病变，尽管在训练阶段未提供此类信息。最后，还介绍和分析了输入图像分辨率和不同 DL 模型架构对诊断准确性的影响。