Institute of Biomedical Engineering, School of Communication and Information Engineering, Shanghai University, Shanghai, China.
University of Science and Technology of China, Hefei, China.
Biomed Eng Online. 2021 Jul 28;20(1):71. doi: 10.1186/s12938-021-00908-1.
The classification of benign and malignant microcalcification clusters (MCs) is an important task for computer-aided diagnosis (CAD) of digital breast tomosynthesis (DBT) images. Influenced by imaging method, DBT has the characteristic of anisotropic resolution, in which the resolution of intra-slice and inter-slice is quite different. In addition, the sharpness of MCs in different slices of DBT is quite different, among which the clearest slice is called focus slice. These characteristics limit the performance of CAD algorithms based on standard 3D convolution neural network (CNN).
To make full use of the characteristics of the DBT, we proposed a new ensemble CNN, which consists of the 2D ResNet34 and the anisotropic 3D ResNet to extract the 2D focus slice features and 3D contextual features of MCs, respectively. Moreover, the anisotropic 3D convolution is used to build 3D ResNet to avoid the influence of DBT anisotropy.
The proposed method was evaluated on 495 MCs in DBT images of 275 patients, which are collected from our collaborative hospital. The area under the curve (AUC) of receiver operating characteristic (ROC) and accuracy of classifying benign and malignant MCs using decision-level ensemble strategy were 0.8837 and 82.00%, which were significantly higher than the experimental results of 2D ResNet34 (AUC: 0.8264, ACC: 76.00%) and anisotropic 3D ResNet (AUC: 0.8455, ACC: 76.00%). Compared with the results of 3D features classification in the radiomics, the AUC of the deep learning method with decision-level ensemble strategy was improved by 0.0435, and the F1 score was improved from 79.37 to 85.71%. More importantly, the sensitivity increased from 78.13 to 84.38%, and the specificity increased from 66.67 to 77.78%, which effectively reduced the false positives of diagnosis CONCLUSION: The results fully prove that the ensemble CNN can effectively integrate 2D features and 3D features, improve the classification performance of benign and malignant MCs in DBT, and reduce the false positives.
良性和恶性微钙化簇（MCs）的分类是数字乳腺断层合成（DBT）图像计算机辅助诊断（CAD）的重要任务。受成像方法的影响，DBT 具有各向异性分辨率的特点，其中层内和层间的分辨率有很大的不同。此外，DBT 中 MCs 的清晰度在不同切片中差异很大，其中最清晰的切片称为焦点切片。这些特点限制了基于标准 3D 卷积神经网络（CNN）的 CAD 算法的性能。
为了充分利用 DBT 的特点，我们提出了一种新的集成 CNN，它由 2D ResNet34 和各向异性 3D ResNet 组成，分别提取 MCs 的 2D 焦点切片特征和 3D 上下文特征。此外，使用各向异性 3D 卷积来构建 3D ResNet，以避免 DBT 各向异性的影响。
该方法在我们合作医院采集的 275 名患者的 495 个 MCs 的 DBT 图像上进行了评估。使用决策级集成策略对良性和恶性 MCs 进行分类的接收者操作特征（ROC）曲线下面积（AUC）和准确率分别为 0.8837 和 82.00%，明显高于 2D ResNet34（AUC：0.8264，ACC：76.00%）和各向异性 3D ResNet（AUC：0.8455，ACC：76.00%）的实验结果。与放射组学中 3D 特征分类的结果相比，决策级集成策略的深度学习方法的 AUC 提高了 0.0435，F1 评分从 79.37%提高到 85.71%。更重要的是，敏感性从 78.13%提高到 84.38%，特异性从 66.67%提高到 77.78%，有效地减少了诊断的假阳性。
结果充分证明了集成 CNN 可以有效地整合 2D 特征和 3D 特征，提高 DBT 中良性和恶性 MCs 的分类性能，并减少假阳性。
