Related Experiment Video
Updated: Jun 20, 2026
10:37
A Multimodal Imaging Framework to Advance Phenotyping of Living Label-free Breast Cancer Cells
Published on: August 22, 2025
Multimodal feature fusion model for breast mass malignant risk stratification.
Shengxin Pei1,2, Xiumei Tang3,4, Hongxia Su5,6
1Department of Ultrasound, West China Hospital/West China School of Medicine, Sichuan University, Chengdu, China.
Frontiers in Oncology
|June 19, 2026
Summary
Machine learning models integrating multimodal breast ultrasound features improve malignancy risk stratification. The Random Forest model showed superior performance, particularly in BI-RADS categories 2, 3, 5, and 4a.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Oncology
Background:
- Accurate breast mass malignancy risk stratification is crucial for patient management.
- Integrating diverse data sources can enhance diagnostic performance.
Purpose of the Study:
- To develop and validate machine learning models for breast mass malignancy risk stratification.
- To compare the diagnostic performance of models using multimodal features (BI-RADS terminology, ultrasound imaging, radiomics) across BI-RADS categories.
Main Methods:
- Retrospective analysis of 3,703 breast ultrasound images from 2,685 patients.
- Training and validation of Logistic Regression, Support Vector Machine, and Random Forest models.
- Evaluation using diagnostic accuracy and AUC, overall and within BI-RADS subcategories.
Main Results:
- The Random Forest model with combined multimodal features achieved the highest overall AUC (0.850).
- Single-modality models showed varying optimal performance (BI-RADS terminology: LR, AUC=0.820; radiomics: LR, AUC=0.740; imaging: RF, AUC=0.800).
- High performance in BI-RADS categories 2, 3, 5, and 4a; lower performance in 4b and 4c.
Conclusions:
- Multimodal machine learning models effectively stratify breast mass malignancy risk.
- The Random Forest model integrating combined features demonstrates superior diagnostic performance.
- Clinical utility is suggested for BI-RADS categories 2, 3, 5, and 4a, with a need for further refinement in categories 4b and 4c.