A Deep Learning Model for Predicting Molecular Subtype of Breast Cancer by Fusing Multiple Sequences of DCE-MRI From...

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate breast cancer molecular subtyping is crucial for treatment selection.
• Dynamic contrast-enhanced MRI (DCE-MRI) offers valuable imaging biomarkers.
• Deep learning (DL) shows promise in analyzing complex medical imaging data.

Purpose of the Study:

• To evaluate the performance of DL in predicting breast cancer molecular subtypes using DCE-MRI.
• To develop and assess a novel multi-branch convolutional neural network (MBCNN) for this task.
• To compare MBCNN with existing DL models.

Main Methods:

• A retrospective study of 366 breast cancer patients from two institutes.
• Development of a multi-branch convolutional neural network (MBCNN) with appearance transformation.
• Assessment of MBCNN using different regions of interest (ROIs) and fusion strategies.

Main Results:

• MBCNN achieved optimal performance with intermediate fusion and an ROI size of 80 pixels.
• MBCNN outperformed CNN and CLSTM in predicting luminal B, HER2-enriched, and TN subtypes.
• For four-subtype prediction, MBCNN achieved an accuracy of 0.64, outperforming CNN (0.48) and CLSTM (0.52).

Conclusions:

• A developed DL model using feature extraction and fusion of DCE-MRI enables preoperative prediction of breast cancer molecular subtypes.
• The MBCNN demonstrated high diagnostic performance for breast cancer molecular subtype prediction.
• This approach facilitates personalized treatment strategies through accurate preoperative subtyping.