Machine and Deep Learning on Radiomic Features from Contrast-Enhanced Mammography and Dynamic Contrast-Enhanced...

Area of Science:

• Medical Imaging
• Machine Learning
• Oncology

Background:

• Accurate breast cancer characterization is crucial for effective treatment planning.
• Radiomics analysis of medical images offers a non-invasive approach to assess tumor characteristics.
• Integrating Dynamic Contrast Enhanced Magnetic Resonance Imaging (DCE-MRI) and contrast-enhanced mammography (CEM) may improve diagnostic accuracy.

Purpose of the Study:

• To evaluate machine and deep learning models using radiomics features from DCE-MRI and CEM.
• To assess the accuracy of these models in characterizing breast cancer.
• To predict tumor molecular profile, including histological grade and HER2 status.

Main Methods:

• 153 patients with malignant and benign lesions underwent DCE-MRI.
• Radiomic features (n=851) were extracted using PyRadiomics, followed by feature selection and normalization.
• Machine learning models (Gradient Boosting Machine, LASSO, Neural Network) were trained and evaluated using cross-validation and AUC metrics.

Main Results:

• Radiomic features from CEM and DCE-MRI demonstrated strong discriminative performance for malignancy (AUCs > 0.80).
• Gradient Boosting Machine achieved the highest accuracy (0.911) for differentiating benign from malignant lesions.
• Neural networks showed promise in predicting tumor grade (accuracy 0.848) and HER2+ status (AUC 0.669).

Conclusions:

• Machine learning models utilizing radiomics from multiparametric CEM and DCE-MRI are promising non-invasive tools for breast cancer characterization.
• These models effectively distinguish benign from malignant lesions and show potential for predicting histological grade and HER2 status.
• Radiomics-based models can aid in early diagnosis, grading, and biomarker assessment, supporting personalized treatment and non-invasive decision-making.