Optimizing S-detect classification accuracy for BI-RADS 4 breast nodules using multimodal ultrasound parameters.

Area of Science:

• Medical imaging analysis
• Artificial intelligence in diagnostics
• Breast cancer detection

Background:

• Deep learning (DL) tool S-detect shows limited diagnostic specificity (59.57%) for Breast Imaging Reporting and Data System (BI-RADS 4) breast lesions.
• The potential of quantitative multimodal ultrasound (MUS) parameters to enhance S-detect's performance is not well-established.

Purpose of the Study:

• To improve the diagnostic accuracy of S-detect for differentiating benign from malignant breast nodules.
• To investigate the effectiveness of integrating MUS parameters into the S-detect tool.

Main Methods:

• Retrospective analysis of clinical and ultrasound data from 231 female patients with BI-RADS 4 breast nodules.
• Extraction and analysis of quantitative MUS parameters (e.g., vascular resistance index, calcification, elastic strain ratio, vascularity index).
• Evaluation of S-detect classification performance before and after optimization with MUS parameters using sensitivity, specificity, accuracy, and AUC.

Main Results:

• Malignant nodules exhibited significantly higher elastic strain ratio (SR) and vascularity index (VI) compared to benign nodules.
• Multivariate analysis identified SR, VI, vascular resistance index, calcification, and specific axis plane features as independent predictors of malignancy.
• The combined S-detect + MUS model achieved high diagnostic performance (SE 86.75%, SP 92.31%, AUC 0.93), significantly improving specificity for BI-RADS 4a lesions.

Conclusions:

• Combining S-detect with MUS parameters substantially enhances the accuracy of differential diagnosis for BI-RADS lesions.
• This integrated approach offers a reliable basis for clinical decision-making in breast nodule assessment.
• Further multi-center prospective studies are recommended to validate these findings.