Related Experiment Video
Updated: May 13, 2025
15:48
Tracking the Mammary Architectural Features and Detecting Breast Cancer with Magnetic Resonance Diffusion Tensor Imaging
Published on: December 15, 2014
22.3K
Optimizing Deep Learning Models for Luminal and Nonluminal Breast Cancer Classification Using Multidimensional ROI in
Zhenfeng Huang1,2, Zhikun Qiu3, Shuyuan Chen4
1Department of Thyroid & Breast Surgery, The Fifth Affiliated Hospital, Sun Yat-sen University, Zhuhai, China.
Cancer Medicine
|May 10, 2025
Summary
A 2.5D deep learning model effectively distinguishes breast cancer subtypes using DCE-MRI. This model, incorporating a 4mm peritumoral region, shows promise as a diagnostic tool for luminal versus nonluminal tumors.
Area of Science:
- Radiology
- Oncology
- Artificial Intelligence
Background:
- Deep learning models are increasingly used in medical imaging analysis.
- Previous studies have not comprehensively evaluated the impact of region of interest (ROI) dimensions, peritumoral expansion, and segmentation strategies on deep learning model performance for breast cancer subtyping.
- Distinguishing molecular subtypes of breast cancer is crucial for guiding treatment decisions.
Purpose of the Study:
- To evaluate the performance of multidimensional deep transfer learning models in distinguishing luminal from nonluminal breast cancer subtypes using DCE-MRI.
- To systematically compare the effects of ROI dimensions (2D/2.5D/3D) and peritumoral expansion levels (0-8mm) under different segmentation scenarios (ROI only vs. ROI original).
- To optimize deep learning models for breast cancer subtyping through transfer learning.
Main Methods:
- Retrospective collection of DCE-MRI data from 426 primary invasive breast cancer patients across three cohorts (training, validation 1, validation 2).
- Delineation of ROIs with subsequent expansions of 2, 4, 6, and 8mm.
- Assessment of various deep transfer learning models using ROC curves and decision curve analysis, considering precise segmentation and peritumoral regions.
Main Results:
- The 2.5D deep learning model, using the ROI original segmentation with a 4mm peritumoral expansion, achieved optimal performance.
- This model demonstrated high AUC values across all cohorts: 0.808 (training), 0.766 (validation 1), and 0.799 (validation 2).
Conclusions:
- A 2.5D deep learning model incorporating the three principal slices of the minimum bounding box (ROI original) and a 4mm peritumoral region effectively distinguishes luminal from nonluminal breast cancer.
- This approach shows potential as a non-invasive diagnostic tool for breast cancer molecular subtyping using DCE-MRI.
