Explainable Radiomics-Based Model for Automatic Image Quality Assessment in Breast Cancer DCE MRI Data

Area of Science:

• Radiomics and Medical Imaging Analysis
• Machine Learning in Healthcare
• Biomedical Signal Processing

Background:

• Accurate assessment of medical imaging quality is crucial for reliable diagnosis.
• Breast cancer Dynamic Contrast-Enhanced Magnetic Resonance Imaging (DCE-MRI) quality can significantly impact diagnostic accuracy.
• Existing methods for image quality assessment may lack automation and explainability.

Purpose of the Study:

• To develop an explainable radiomics-based model for automated image quality assessment in breast cancer DCE-MRI.
• To evaluate the performance of machine learning classifiers in distinguishing high-quality from low-quality DCE-MRI scans.
• To enhance the reliability and fairness of large-scale medical imaging datasets through quality control.

Main Methods:

• Extraction of 819 radiomic features and 2 No-Reference image quality metrics from 280 breast cancer DCE-MRI images.
• Feature extraction from the whole image and background regions of interest, considering two scenarios: 12 slices per patient and the middle slice.
• Application of machine learning classifiers (including Support Vector Machine) with explainability assessed using SHapley Additive Explanations (SHAP).

Main Results:

• The model achieved the best performance when using features from the middle slice (scenario ii), combining whole image and background features.
• A Support Vector Machine classifier yielded high performance metrics: 85.51% sensitivity, 80.01% specificity, 82.76% accuracy, and 89.37% AUC.
• The SHAP analysis provided explainability for the model's predictions, identifying key features influencing quality assessment.

Conclusions:

• The developed explainable radiomics model shows significant potential for automatic image quality assessment in breast cancer DCE-MRI.
• The model can be integrated into clinical workflows to ensure data quality and improve diagnostic reliability.
• This approach offers a valuable tool for managing large imaging repositories and conducting fair subgroup analyses.