Longitudinal MRI-based deep learning model for predicting pathological complete response in breast cancer: a...

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Medicine
• Oncology

Background:

• Early prediction of neoadjuvant therapy (NAT) response is vital for tailoring breast cancer treatment.
• Current methods may lack the ability to fully utilize longitudinal data for dynamic tumor change assessment.

Purpose of the Study:

• To develop and validate a novel breast self-supervised temporal learning framework (BSTNet) for predicting pathological complete response (pCR) to NAT.
• To assess BSTNet's generalizability and performance across multi-center cohorts and diverse molecular subtypes.

Main Methods:

• Utilized a self-supervised temporal learning framework (BSTNet) on longitudinal MRI data from 1339 patients.
• Employed multi-center internal and external validation strategies.
• Conducted subgroup analyses based on molecular subtypes and MRI timing.

Main Results:

• BSTNet achieved high predictive performance with AUCs of 0.882 (internal) and 0.857-0.854 (external validation).
• Consistent performance was observed across different molecular subtypes (AUCs 0.818-0.895).
• The model demonstrated robust specificity in identifying non-pCR patients (74.5%-86.4%).

Conclusions:

• BSTNet offers a robust and generalizable deep learning solution for early pCR prediction in breast cancer.
• The framework's ability to interpret variable longitudinal MRI data supports adaptive treatment planning in clinical practice.