Deep-learning-based HER2 status assessment from multimodal breast cancer data predicts neoadjuvant therapy response

Area of Science:

• Oncology
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate human epidermal growth factor receptor 2 (HER2) status assessment is vital for breast cancer treatment.
• Intratumoural heterogeneity limits the accuracy of traditional needle biopsies.

Purpose of the Study:

• To introduce a deep-learning model for accurate HER2 status prediction using pretreatment multimodal breast cancer images.
• To compare the HER2 prediction performance of the model against needle biopsies in patients undergoing neoadjuvant therapy.

Main Methods:

• Development of the HER2 multimodal alignment and prediction (MAP) model using deep learning.
• Leveraging pretreatment multimodal breast cancer images for comprehensive tumor characteristic reflection.
• Validation on a large-scale dataset of 14,472 images from 6,991 breast cancer cases across 4 centers.

Main Results:

• The HER2 MAP model demonstrated superior performance in predicting patient response compared to needle biopsies.
• Consistent superiority of the HER2 MAP model was observed across the large-scale dataset.
• The model provides more accurate HER2 status prediction, reflecting a comprehensive view of tumor characteristics.

Conclusions:

• The deep-learning-based HER2 MAP model offers a significant advancement over traditional methods for HER2 status assessment.
• This tool aids physicians in making informed clinical decisions for breast cancer treatment planning.
• Improved HER2 prediction accuracy has the potential to enhance patient outcomes in breast cancer care.