# Dynamic Contrast-Enhanced MRI Kinetic Curve-Driven Parametric Radiomics for Predicting Breast Cancer Molecular Subtypes: A Multicenter and Interpretable Study

## TL;DR

This study introduces a non-invasive method using MRI data to accurately predict breast cancer subtypes, aiding in personalized treatment.

## Contribution

The novel approach uses DCE-MRI kinetic curves to create interpretable radiomics models for breast cancer subtype prediction.

## Key findings

- The TIC-Combined model achieved high predictive performance with AUCs of 0.79 and 0.77 in internal and external validation sets.
- The model showed strong subtype-specific classification, particularly for triple-negative and HER2+ breast cancers.
- The method demonstrated good calibration and high interpretability, supporting reliable clinical predictions.

## Abstract

Reliable prediction of breast cancer molecular subtypes is critical for guiding personalized treatment and improving clinical outcomes. Our study proposes an innovative, non-invasive parametric radiomics approach derived from DCE-MRI time-intensity curve kinetics. By converting original multiphase images into parametric images, and applying advanced radiomics and machine learning methods, we developed and validated interpretable models capable of accurately classifying breast cancer molecular subtypes. The findings in this study highlight the potential of DCE-MRI kinetic-driven radiomics to provide clinically meaningful, non-invasive subtype prediction, thereby supporting precision oncology.

Background/Objectives: To investigate and develop a non-invasive parametric radiomics model derived from dynamic contrast-enhanced MRI (DCE-MRI) time-intensity curve (TIC) kinetics for predicting breast cancer molecular subtypes (HR+/HER2−, HER2+ and triple-negative breast cancer). Methods: This multicenter retrospective study enrolled 935 female patients with histologically confirmed breast cancer who underwent pretreatment breast DCE-MRI from August 2017 to July 2022. Based on the wash-in rate (WIR) and the area under the TIC, the original multiphase DCE-MRI images were converted into two types of parametric images. Radiomics features were extracted from TIC-WIR and TIC-Area images and analyzed using low variance filtering, the elimination of highly correlated features, and the least absolute shrinkage and selection operator regression. The categorical boosting algorithm was employed to develop multiclass prediction models for breast cancer molecular subtyping. A TIC-Combined model was further established by integrating the calibrated probability outputs of the TIC-WIR and TIC-Area models using a decision-level fusion strategy. The discrimination, calibration, and interpretability of the models were evaluated in the study datasets. Results: The TIC-Combined model achieved superior predictive performance in both the internal validation set (micro-average AUC: 0.79, macro-average AUC: 0.77) and the external validation set (micro-average AUC: 0.77, macro-average AUC: 0.75). For subtype-specific classification by the TIC-Combined model, the highest one-vs-rest AUCs were 0.81 for triple-negative breast cancer in the internal validation set and 0.76 for HER2+ breast cancer in the external validation set. The TIC-Combined model also showed good calibration and high interpretability which ensured reliable predictions and provided clear insights into feature importance. Conclusions: Interpretable parametric radiomics from TIC-derived parametric maps links kinetic features to molecular phenotypes, enabling accurate and non-invasive classification of breast cancer molecular subtypes.