Deep Radiogenomics Sequencing for Breast Tumor Gene-Phenotype Decoding Using Dynamic Contrast Magnetic Resonance...

Area of Science:

• Radiogenomics
• Medical Imaging
• Machine Learning

Background:

• Accurate prediction of Estrogen Receptor (ER), Progesterone Receptor (PR), and Human Epidermal Growth Factor Receptor 2 (HER2) gene status is crucial for breast cancer treatment decisions.
• Current methods often rely on invasive biopsies, highlighting the need for non-invasive predictive techniques.
• Dynamic contrast-enhanced magnetic resonance imaging (MRI) offers rich spatial and temporal information that may correlate with tumor genetic profiles.

Purpose of the Study:

• To investigate the feasibility of radiogenomic profiling for predicting ER, PR, and HER2 gene status in breast cancer patients using dynamic contrast-enhanced MRI.
• To evaluate the performance of deep learning algorithms in decoding gene-phenotype relationships from breast MRI data.

Main Methods:

• Utilized a dataset of 922 invasive breast cancer patients with known ER, PR, and HER2 mutation status.
• Employed 3D deep learning networks analyzing T1-weighted pre-contrast and three post-contrast MRI sequences for multi-channel analysis.
• Implemented N4 bias correction and standardized input sizes (128x128x68) for all networks, with data split into training/validation (80%) and testing (20%) sets.

Main Results:

• For ER prediction, SEResNet50 achieved an Area Under the Curve (AUC) of 0.695 (sensitivity: 0.564, specificity: 0.787).
• For PR prediction, ResNet34 achieved an AUC of 0.658 (sensitivity: 0.593, specificity: 0.734).
• For HER2 prediction, SEResNext101 achieved an AUC of 0.698 (sensitivity: 0.750, specificity: 0.625).

Conclusions:

• The study demonstrates the potential of using deep learning algorithms on dynamic contrast-enhanced MRI for non-invasive radiogenomic profiling of breast tumors.
• The developed models achieved moderate performance in predicting ER, PR, and HER2 gene statuses, indicating the feasibility of imaging-based gene-phenotype decoding.