基于医学显著特征的图神经网络用于乳腺超声恶性检测
该研究在 BUS、BUS2 及 Mendeley 乳腺超声数据集中提取形状、纹理与统计学等医学相关特征,以图节点和 k 近邻边构建图结构,采用 GraphSAGE 分类器进行良恶性判别。结果显示 LSTM 聚合器在 BUS 达 94.13%、Mendeley 达 99.99%、BUS2 达 73.19% 的准确率。
要点
- 模型融入形状、纹理及统计特征,不只依赖原始图像像素,更贴近影像科医生诊断逻辑。
- LSTM 聚合策略在多个数据集中总体优于均值与最大池化聚合,是值得关注的建模选择。
- 不同数据集中准确率差异较大(如 BUS2 仅约 73%),外部推广时应审慎验证。
核心结论
- GraphSAGE 图神经网络整合医学显著特征后可有效检测乳腺超声恶性病变,并在多个数据集间展现出较强泛化能力。
- 图结构通过特征相似性建模图像间关系,能捕捉病变之间的相似性并提升分类性能。
- 该方法为利用医学意义描述符改进乳腺超声计算机辅助诊断提供了新思路。
研究设计与样本
- 研究采用 BUS、BUS2 和 Mendeley Breast Ultrasound 三个乳腺超声数据集,具体样本量及中心数原文未提供。
- 预处理中通过按位与运算进行区域提取,以定位病变相关区域。
- 特征提取采用方向梯度直方图、灰度共生矩阵及直方图描述符,获取形状、纹理与统计学特征。
- 每幅图像表示为一个带特征属性的图节点,并基于特征相似度通过 k 近邻方法构建边。
- 分类器使用 GraphSAGE 图神经网络,聚合器包括均值、最大值及 LSTM 三种策略。
关键影像表现/征象
- 研究关注的特征为肿瘤形态、纹理及统计学属性,如通过灰度共生矩阵提取的纹理特征。
- 图结构中以特征相似度相连的节点,其影像特征具有可比性,据此捕捉病变间相似关系。
诊断标准与关键数值
- BUS 数据集上准确率:均值聚合 92.89%,最大聚合 93.20%,LSTM 聚合 94.13%。
- Mendeley Breast Ultrasound 数据集准确率:均值 99.80%,最大 99.60%,LSTM 99.99%。
- BUS2 数据集准确率:均值 71.96%,最大 73.15%,LSTM 73.19%。
- 合并 BUS 与 BUS2 数据集准确率:均值 86.69%,最大 86.58%,LSTM 87.62%。
- 原文未报告敏感度、特异度、AUC 及具体分类阈值。
临床应用与阅片要点
- 该框架提示在计算机辅助诊断中融入形状和纹理等医学特征可提高分类性能,为开发更贴近临床的模型提供参考。
- LSTM 聚合策略在多数数据集中表现较优,可作为后续模型优化时优先考虑的聚合方式。
- BUS2 数据集准确率明显低于其他数据集,提示在不同设备和采集条件下应用时需注意模型表现下降。
局限性
- 原文未提供样本量、中心数及参考标准等细节,读者难以评估结果的可重复性和临床适用性。
- BUS2 数据集上的准确率明显偏低,模型在该数据集上的泛化能力受限。
- 研究仅报告准确率,缺少敏感度、特异度及受试者工作特征曲线等更全面的诊断性能指标。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
英文原文(摘要)· 点击展开
OBJECTIVE: Breast ultrasound imaging is widely used for the early detection of malignant breast lesions. Although deep learning models have shown strong performance, most approaches rely solely on raw image features and overlook clinically significant characteristics essential for reliable diagnosis. This study aims to develop a graph-based neural network model that integrates medically meaningful features to improve malignancy detection in breast ultrasound images.
METHODS: The Breast Ultrasound (BUS) and Breast Ultrasound 2 (BUS2) datasets were pre-processed for region extraction using a bit-wise AND operation. Clinically relevant tumor features, such as shape, texture and statistical attributes, were extracted using Histogram of Oriented Gradients, gray-level co-occurrence matrix (GLCM) and histogram descriptors from the BUS, BUS2 and Mendeley Breast Ultrasound datasets. Each image was represented as a graph node with feature attributes, with the edges constructed using k-nearest neighbors based on feature similarity. A GraphSAGE-based graph neural network classifier with mean, max and long short-term memory (LSTM) aggregators was trained and evaluated.
RESULTS: On the BUS dataset, the proposed approach achieved accuracies of 92.89% (mean), 93.20% (max) and 94.13% (LSTM). On the Mendeley Breast Ultrasound dataset, accuracies reached 99.80% (mean), 99.60% (max) and 99.99% (LSTM). For the BUS2 dataset, accuracies were 71.96% (mean), 73.15% (max) and 73.19% (LSTM). On the combined BUS-BUS2 dataset, the model achieved 86.69% (mean), 86.58% (max) and 87.62% (LSTM).
CONCLUSION: The proposed GraphSAGE-based graph neural network effectively incorporates clinically significant features for breast ultrasound malignancy detection and demonstrates strong generalization across multiple datasets. By modeling feature-based relationships among ultrasound images through a graph structure, the proposed framework captures similarities between lesions and enhances classification performance. These findings highlight the potential of graph-based representations for improving computer-aided breast cancer diagnosis using medically meaningful descriptors.
引用
- Deb, Dash, Mohapatra, et al. Graph-Based Neural Network for Malignancy Detection in Breast Ultrasound Images using Medically Significant Features. Ultrasound Med Biol. 2026. PMID: 42674880. doi:10.1016/j.ultrasmedbio.2026.07.029
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。