Explainable breast cancer molecular expression prediction using multi-task deep-learning based on 3D whole breast...

Area of Science:

• Oncology
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate prediction of breast cancer biomarkers (ER, PR, HER2) is crucial for treatment decisions.
• Current methods often require invasive procedures, highlighting the need for noninvasive estimation.
• Deep learning offers potential for enhancing diagnostic capabilities in medical imaging.

Purpose of the Study:

• To noninvasively estimate breast cancer biomarkers ER, PR, and HER2 using 3D whole breast ultrasound (3DWBUS).
• To enhance the performance and interpretability of biomarker prediction models through multi-task deep learning.
• To compare a multi-task model against a single-task model for predicting these biomarkers.

Main Methods:

• 388 breast cancer patients underwent 3DWBUS examinations.
• Two deep learning models were developed: a single-task model for biomarker prediction and a multi-task model combining tumor segmentation with biomarker prediction.
• Model performance was evaluated using AUC metrics and compared with Delong's test; interpretability was assessed via Grad-CAM++ visualization.

Main Results:

• The multi-task model demonstrated superior overall performance in the test set, achieving a higher macro AUC (0.733) compared to the single-task model (0.708).
• Individual biomarker prediction showed varied performance, with the multi-task model outperforming the single-task model for PR and HER2.
• Grad-CAM++ visualization indicated that the multi-task model focused more effectively on diseased tissue, enhancing interpretability.

Conclusions:

• The multi-task deep learning model effectively predicts breast cancer biomarkers noninvasively from 3DWBUS images.
• The multi-task approach improves prediction accuracy and offers enhanced clinical interpretability compared to single-task models.
• This technology holds promise for improving targeted drug screening efficiency and clinical decision-making in breast cancer management.