Graph neural network-based multi-label classification of breast cancer using multiparametric MRI and clinical features

基于图神经网络的乳腺癌多标签分类,利用多参数磁共振成像和临床特征

阅读:3

Abstract

This research was done to design and evaluate a graph-based deep-learning architecture to automatically multi-label breast cancer, using multiparametric magnetic resonance imaging (MRI) data and clinical features. The general aim was to enhance the diagnostic accuracy of the key clinical activities, i.e. biomarker discovery, tumour staging, and histological grading. Three-dimensional volumetric graphs were built with the help of five different MRI modalities (Phase 1, Phase 2, Phase 3, T1, and IHSFused), and each voxel and clinical variable was a node. Attributes of the nodes were radiomic features and patient-specific clinical data whereas edges represented relationships by spatial proximity and intensity. Graph structure processing was done using a Graph Isomorphism Network (GIN), which gradually optimised node embeddings by aggregating neighbourhoods. The multi-label classification focused on the HER2 status, TNM stage (Tumour Size, Node involvement, Metastasis) and histological grade elements (Tubule formation, Mitotic count, Nuclear pleomorphism). Cross-Entropy loss and Adam optimiser were used to train the model, and Synthetic Minority Over-Sampling Technique (SMOTE) was used to balance the distribution of classes. The protocol of cross-validation was used five times to be robust and performance measures were provided in terms of accuracy, precision, recall, F1-score, and confusion matrices. The t-Stochastic Neighbor Embedding (t-SNE) visualisations were used to assess the feature separability. The model achieved high accuracy in HER2 (up to 91.42%) and Tubule grade (up to 91.83%), moderate performance in Tumor Size (70–72%), Nodes (82–84%), and Metastasis (80–83%), and lower accuracy in nuclear grade (64–68%). Confusion matrices showed minimal misclassification in high-performing labels and greater variability in lower-performing ones. t-SNE plots confirmed effective feature separation for HER2 and Tubule, with overlap noted in Nuclear. These results indicate the possibilities of volumetric graph-based learning to provide full classification of breast cancer and its importance in improving precision diagnostics with the help of artificial intelligence. SUPPLEMENTARY INFORMATION: The online version contains supplementary material available at 10.1007/s12672-025-04359-1.

特别声明

1、本页面内容包含部分的内容是基于公开信息的合理引用;引用内容仅为补充信息,不代表本站立场。

2、若认为本页面引用内容涉及侵权,请及时与本站联系,我们将第一时间处理。

3、其他媒体/个人如需使用本页面原创内容,需注明“来源:[生知库]”并获得授权;使用引用内容的,需自行联系原作者获得许可。

4、投稿及合作请联系:info@biocloudy.com。