Development and validation of an MRI-based deep learning system for triple-class ER expression classification in...

Area of Science:

• Oncology
• Radiology
• Artificial Intelligence

Background:

• Estrogen receptor (ER) expression is crucial for breast cancer prognosis and predicting response to endocrine therapy (ET).
• The 2020 ASCO/CAP guidelines define ER-positive as ≥1% ER-positive nuclei, but ER-low positive (1%-10%) breast cancer has distinct biology and uncertain ET response.
• Invasive ER assessment has limitations, necessitating non-invasive methods for accurate classification.

Purpose of the Study:

• To develop and validate a non-invasive, MRI-based deep learning system (BERC) for classifying ER negative, ER-low positive, and ER-high positive breast cancer.
• To assess the performance of the BERC system using multicenter data.

Main Methods:

• Retrospective analysis of pretreatment DCE-MRI data from 3500 breast cancer patients across six institutions.
• Automated deep learning for tumor segmentation and development of a DCE-MRI-based classification model.
• Evaluation of model performance using AUC, sensitivity, specificity, accuracy, PPV, and NPV; interpretability assessed via t-SNE, UMAP, and SHAP analysis.

Main Results:

• The model achieved high performance in training (Micro-AUC 0.918, Macro-AUC 0.882) and internal validation (Micro-AUC 0.923, Macro-AUC 0.900).
• External testing across four centers demonstrated robust performance with Micro-average AUCs ranging from 0.828 to 0.923 and Macro-average AUCs from 0.825 to 0.905.
• Feature visualization confirmed biologically plausible clustering patterns for ER expression categories.

Conclusions:

• The BERC system demonstrates significant potential for non-invasive, preoperative prediction of ER status in breast cancer.
• This AI-driven approach may facilitate personalized treatment strategies, particularly for patients with ER-low positive disease.