Radiomics-Based Machine Learning Models for Classifying Breast Cancer on Dynamic-Contrast Enhanced MRI through...

Area of Science:

• Radiology
• Medical Imaging
• Machine Learning

Background:

• Breast cancer diagnosis requires efficient methods to improve patient survival rates.
• Dynamic Contrast-Enhanced MRI (DCE-MRI) is a key imaging modality for breast cancer assessment.
• Manual segmentation of DCE-MRI images can be time-consuming and subject to inter-observer variability.

Purpose of the Study:

• To evaluate observer performance in manual segmentation of DCE-MRI.
• To assess the efficacy of radiomic features and machine learning (ML) for classifying breast lesions.
• To identify key radiomic features predictive of malignancy.

Main Methods:

• Radiologists manually segmented 155 breast lesions (65 benign, 90 malignant) on DCE-MRI images.
• 107 radiomic features were extracted and normalized; feature selection was performed using LASSO regression.
• Nine ML models were trained and evaluated, with CatBoost showing superior performance; SHAP analysis identified key features.

Main Results:

• Excellent inter-rater reliability (ICC: 0.941-0.992) was observed for radiomic feature extraction.
• The CatBoost model achieved the highest AUC of 0.937, with sensitivity of 0.889 and specificity of 0.909.
• SHAP analysis confirmed the significance of specific radiomic features in differentiating benign from malignant lesions.

Conclusions:

• Radiomic features combined with ML, particularly ensemble methods like CatBoost, offer high potential for accurate breast cancer classification.
• The study highlights the effectiveness of CatBoost in distinguishing malignant from benign breast lesions using DCE-MRI data.
• Interpretable AI models, like CatBoost with SHAP analysis, can enhance diagnostic confidence and clinical utility.