Transformer-Based Deep Learning Model Using MRI-Derived Microvascular Atlas for Predicting Lymphovascular Invasion in...

Area of Science:

• Radiology
• Oncology
• Artificial Intelligence

Background:

• Lymphovascular invasion (LVI) is a critical prognostic marker in breast cancer, associated with metastasis and poor outcomes.
• Accurate preoperative identification of LVI is essential for guiding treatment strategies.

Purpose of the Study:

• To develop a noninvasive deep learning model for predicting LVI status in breast cancer patients.
• To integrate MRI-derived microvascular parameters with a deep learning architecture for enhanced prediction.
• To explore the biological interpretability of the model's predictions.

Main Methods:

• Retrospective analysis of 436 breast cancer patients with pathologically confirmed LVI status.
• High-resolution multi-b-value diffusion-weighted imaging (DWI) to derive microvascular parameters (ADC, v_m, v_s, ANB).
• A TwinsSVT deep learning model incorporating a multi-parametric fusion module for LVI prediction.

Main Results:

• The multi-parametric fusion model achieved high predictive performance (AUCs of 0.881 and 0.859 in internal and external validation).
• Grad-CAM visualizations indicated the model focused on tumor margins and vascular density, correlating with pathological findings.
• The model significantly outperformed single-parameter prediction models.

Conclusions:

• A novel deep learning model based on MRI microvascular atlases enables noninvasive preoperative prediction of LVI in breast cancer.
• The model offers biological interpretability, highlighting tumor margins and vascularity.
• This imaging-based tool shows potential for precision subtyping and clinical decision support in breast cancer management.