Scholarlink

This study aims to explore the impact of kinetic enhancement patterns for automated segmentation of dynamic contrast enhanced (DCE) magnetic resonance imaging (MRI) breast cancer data in order to predict pathologic complete response to neoadjuvant chemotherapy (pCR). To this end, a publically available dataset (Duke-Breast-Cancer-MRI) was used including 251 patients. A heuristic scheme was introduced to (a) classify each DCE curve over time in the tumor on a voxel basis into 5 categories including type 1 (plateau), 2 (slow washout), 3 (fast washout), 4 (slow persistent enhancement) and 5 (fast persistent enhancement); and (b) to produce accurate regions of interest (ROIs) for the automated prediction of pCR through radiomics using a variety of machine learning (ML) classifiers. ML radiomics were also investigated using the original tumor’s annotation (i.e., bounding boxes). The kinetic analysis showed that 60% of the tumor is dominantly described by continuous enhancement voxels, and the performance of pCR classification exhibited a maximum ACC of 57% with both automated and original ROIs.

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

Radiomics Prediction of Pathologic Complete Response to Neoadjuvant Chemotherapy in Breast Cancer: Interpretation and Imaging Pitfalls

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Georgios S. Ioannidis,
  • Smriti Joshi,
  • Grigorios Kalliatakis,
  • Katerina Nikiforaki,
  • Vassilis Kilintzis,
  • Haridimos Kondylakis,
  • Oliver Diaz,
  • Maciej Bobowicz,
  • Karim Lekadir,
  • Kostas Marias

摘要

This study aims to explore the impact of kinetic enhancement patterns for automated segmentation of dynamic contrast enhanced (DCE) magnetic resonance imaging (MRI) breast cancer data in order to predict pathologic complete response to neoadjuvant chemotherapy (pCR). To this end, a publically available dataset (Duke-Breast-Cancer-MRI) was used including 251 patients. A heuristic scheme was introduced to (a) classify each DCE curve over time in the tumor on a voxel basis into 5 categories including type 1 (plateau), 2 (slow washout), 3 (fast washout), 4 (slow persistent enhancement) and 5 (fast persistent enhancement); and (b) to produce accurate regions of interest (ROIs) for the automated prediction of pCR through radiomics using a variety of machine learning (ML) classifiers. ML radiomics were also investigated using the original tumor’s annotation (i.e., bounding boxes). The kinetic analysis showed that 60% of the tumor is dominantly described by continuous enhancement voxels, and the performance of pCR classification exhibited a maximum ACC of 57% with both automated and original ROIs.