Predicting pathological complete response based on weakly and semi-supervised joint learning from breast cancer MRI

Area of Science:

• Oncology
• Medical Imaging
• Artificial Intelligence

Background:

• Neoadjuvant chemotherapy (NAC) is standard for breast cancer, with pathological complete response (pCR) indicating a good prognosis.
• Accurate prediction of pCR remains challenging due to limitations in 3D MRI annotation and integrating diverse data.
• Early prediction of tumor response to NAC is vital for treatment stratification.

Purpose of the Study:

• To develop a robust method for predicting pCR in breast cancer patients receiving NAC.
• To integrate multi-parametric MRI attentional features with radiomic features for enhanced prediction accuracy.
• To address the challenge of unifying annotation information for early tumor response prediction.

Main Methods:

• A weakly and semi-supervised joint learning framework was proposed.
• Attention-based multi-instance learning (MIL) was utilized to extract informative MRI features and identify key instances.
• A mean-teacher framework facilitated semi-supervised tumor segmentation for radiomic feature extraction.

Main Results:

• The proposed method achieved an Area Under the Curve (AUC) of 0.85 for pCR prediction.
• This performance significantly outperformed comparative methods in predicting pCR.
• Multi-parametric MRI data demonstrated superior predictive power compared to single-parameter MRI.

Conclusions:

• The developed weakly and semi-supervised learning approach effectively predicts pCR in breast cancer patients undergoing NAC.
• Integrating multi-parametric MRI and radiomic features offers a promising strategy for early tumor response assessment.
• This method holds potential for improving treatment decisions and patient outcomes in breast cancer management.