Keck School of Medicine, University of Southern California, Los Angeles, CA, 90033, USA.
Department of Biomedical Engineering and NIH Resource Center for Medical Ultrasonic Transducer Technology, University of Southern California, Los Angeles, CA, 90089, USA.
Eur Radiol. 2020 May;30(5):3023-3033. doi: 10.1007/s00330-019-06610-0. Epub 2020 Jan 31.
To develop a dual-modal neural network model to characterize ultrasound (US) images of breast masses.
A combined US B-mode and color Doppler neural network model was developed to classify US images of the breast. Three datasets with breast masses were originally detected and interpreted by 20 experienced radiologists according to Breast Imaging-Reporting and Data System (BI-RADS) lexicon ((1) training set, 103212 masses from 45,433 + 12,519 patients. (2) held-out validation set, 2748 masses from 1197 + 395 patients. (3) test set, 605 masses from 337 + 78 patients). The neural network was first trained on training set. Then, the trained model was tested on a held-out validation set to evaluate agreement on BI-RADS category between the model and the radiologists. In addition, the model and a reader study of 10 radiologists were applied to the test set with biopsy-proven results. To evaluate the performance of the model in benign or malignant classifications, the receiver operating characteristic curve, sensitivities, and specificities were compared.
The trained dual-modal model showed favorable agreement with the assessment performed by the radiologists (κ = 0.73; 95% confidence interval, 0.71-0.75) in classifying breast masses into four BI-RADS categories in the validation set. For the binary categorization of benign or malignant breast masses in the test set, the dual-modal model achieved the area under the ROC curve (AUC) of 0.982, while the readers scored an AUC of 0.948 in terms of the ROC convex hull.
The dual-modal model can be used to assess breast masses at a level comparable to that of an experienced radiologist.
• A neural network model based on ultrasonic imaging can classify breast masses into different Breast Imaging-Reporting and Data System categories according to the probability of malignancy. • A combined ultrasonic B-mode and color Doppler neural network model achieved a high level of agreement with the readings of an experienced radiologist and has the potential to automate the routine characterization of breast masses.
开发一种双模态神经网络模型，以对乳腺肿块的超声（US）图像进行特征描述。
本研究开发了一种联合 US 灰阶和彩色多普勒的神经网络模型，用于对乳腺 US 图像进行分类。最初，由 20 位经验丰富的放射科医生根据乳腺影像报告和数据系统（BI-RADS）词汇表对乳腺肿块（（1）训练集，45433 名+12519 名患者的 103212 个肿块。（2）验证集，1197 名+395 名患者的 2748 个肿块。（3）测试集，337 名+78 名患者的 605 个肿块）进行检测和解释。神经网络首先在训练集上进行训练。然后，将训练好的模型在验证集上进行测试，以评估模型和放射科医生对 BI-RADS 类别的分类一致性。此外，将模型和 10 位放射科医生的读者研究应用于具有活检结果的测试集。为了评估模型在良性或恶性分类中的性能，比较了受试者工作特征曲线、敏感性和特异性。
在验证集中，经过训练的双模态模型在将乳腺肿块分为四个 BI-RADS 类别方面与放射科医生的评估具有良好的一致性（κ=0.73；95%置信区间，0.71-0.75）。对于测试集中良性或恶性乳腺肿块的二分类，双模态模型获得的受试者工作特征曲线下面积（AUC）为 0.982，而读者在 AUC 方面的表现为 0.948 基于 ROC 凸壳。
双模态模型可用于评估乳腺肿块，其水平可与经验丰富的放射科医生相媲美。
• 一种基于超声成像的神经网络模型可以根据恶性肿瘤的概率将乳腺肿块分为不同的 BI-RADS 类别。• 联合超声灰阶和彩色多普勒的神经网络模型与经验丰富的放射科医生的读数具有高度一致性，并且有可能实现对乳腺肿块的常规特征描述自动化。
