Deep Learning-Based Recurrence Prediction in HER2-Low Breast Cancer: Comparison of MRI-Alone,...

Area of Science:

• Oncology
• Radiology
• Artificial Intelligence

Background:

• HER2-low breast cancer recurrence risk stratification remains challenging.
• Accurate prediction models are crucial for personalized treatment and follow-up strategies.

Purpose of the Study:

• To develop and compare deep learning (DL) models for predicting recurrence risk in HER2-low breast cancer.
• To evaluate the performance of MRI-alone, clinicopathological-alone, and combined DL models.

Main Methods:

• Analysis of 453 HER2-low breast cancer patients with preoperative MRI and clinicopathological data.
• Development of three DL models: MRI-CNN, clinicopathological-MLP, and a combined multimodal model.
• Performance evaluation using AUC, sensitivity, specificity, and F1-score.

Main Results:

• The clinicopathological-alone model showed the highest AUC (0.92) but lower specificity (72.3%).
• The MRI-alone model had an AUC of 0.69.
• The combined model achieved a balanced performance with AUC 0.90, sensitivity 80.0%, specificity 83.2%, and the highest F1-score (0.55).

Conclusions:

• A DL model integrating MRI and clinicopathological features demonstrates superior performance for predicting recurrence in HER2-low breast cancer.
• This multimodal approach provides a framework for individualized risk assessment.
• The findings may help refine follow-up strategies for HER2-low breast cancer patients.