Scholarlink

Accurate prediction of breast cancer recurrence remains a major challenge in precision oncology, primarily due to the biological heterogeneity of tumor subtypes and the limitations of current genomic assays. In this work, we introduce a dual-stream, multimodal framework that combines radiomic features extracted from both whole-breast parenchymal tissue and segmented tumor regions with clinicopathological variables. Leveraging a fully automated pipeline for segmentation and feature extraction, we compute enhancement, symmetry, and texture descriptors from DCE-MRI volumes, alongside curated clinical data. A transformer-based model (TabNet) is employed for recurrence risk prediction, demonstrating superior discriminative performance (AUC of 0.8193, F1 of 0.7143) compared to conventional machine learning approaches. Feature attribution via SHAP confirms the prognostic relevance of bilateral symmetry, tumor texture, and hormone receptor status. These findings underscore the potential of interpretable, multimodal deep learning for non-invasive, scalable, and individualized recurrence risk stratification ( https://github.com/adnankhalid7454/BC_Recurrence_Pred .).

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Towards Breast Cancer Recurrence Prediction Using Transformer-Based Learning from Global–Local Radiomics and Clinical Data

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Adnan Khalid,
  • Muhammad Mursil,
  • Fabrice Meriaudeau,
  • Ibrahima Faye,
  • Alain Lalande,
  • Domenec Puig,
  • Hatem A. Rashwan

摘要

Accurate prediction of breast cancer recurrence remains a major challenge in precision oncology, primarily due to the biological heterogeneity of tumor subtypes and the limitations of current genomic assays. In this work, we introduce a dual-stream, multimodal framework that combines radiomic features extracted from both whole-breast parenchymal tissue and segmented tumor regions with clinicopathological variables. Leveraging a fully automated pipeline for segmentation and feature extraction, we compute enhancement, symmetry, and texture descriptors from DCE-MRI volumes, alongside curated clinical data. A transformer-based model (TabNet) is employed for recurrence risk prediction, demonstrating superior discriminative performance (AUC of 0.8193, F1 of 0.7143) compared to conventional machine learning approaches. Feature attribution via SHAP confirms the prognostic relevance of bilateral symmetry, tumor texture, and hormone receptor status. These findings underscore the potential of interpretable, multimodal deep learning for non-invasive, scalable, and individualized recurrence risk stratification ( https://github.com/adnankhalid7454/BC_Recurrence_Pred .).