Adaptive Machine Learning Approach for Importance Evaluation of Multimodal Breast Cancer Radiomic Features.

Area of Science:

• Oncology
• Radiology
• Medical Imaging

Background:

• Breast cancer is a leading cause of death among women, necessitating advanced diagnostic tools.
• Quantitative analysis of radiological images offers potential for early detection and classification of breast tumors.
• The P.I.N.K study established the first combined Automated Breast Volume Scanner (ABVS) and Digital Breast Tomosynthesis (DBT) dataset.

Purpose of the Study:

• To perform radiomic analysis for malignant versus benign breast cancer classification using paired ABVS and DBT images.
• To compare the performance of ABVS and DBT modalities in breast cancer classification.
• To identify informative radiomic features and assess the predictive capacity of radiomic models.

Main Methods:

• Development of the first ABVS+DBT dataset with annotated cancerous lesions from 66 women.
• Training radiomic models using a leave-one-out nested cross-validation strategy with threshold selection.
• Evaluation of model performance using Area Under the Receiver Operating Characteristic Curve (AUC-ROC) with varying numbers of features.

Main Results:

• Radiomic models demonstrated predictive capacity with a reduced number of features for both DBT and ABVS.
• Digital Breast Tomosynthesis (DBT) showed superior predictive power compared to Automated Breast Volume Scanner (ABVS).
• High concordance between ABVS and DBT predictions at the patient level (8.7% misclassified by both), indicating partial complementarity.
• Promising results achieved using non-geometric features, suggesting potential for virtual biopsy integration.

Conclusions:

• Radiomic analysis using ABVS and DBT datasets is effective for breast cancer classification.
• The complementarity of ABVS and DBT suggests combined use can improve diagnostic accuracy.
• Non-geometric radiomic features show promise for integrating virtual biopsy into routine medical practice.