Predicting Breast Cancer with Super-Resolution Ultrasound-Based Radiomics: A Multicenter Retrospective Study.

RATIONALE AND

OBJECTIVES

Harnessing super-resolution ultrasound (SRUS) for breast lesion characterization has been limited by manual, parameter-based analyses. Microvasculomics offers an automated solution for quantitative microvascular characterization. We aim to develop and validate a radiomic framework integrating B-mode ultrasound (US), SRUS, and clinical parameters for distinguishing benign from malignant breast lesions.

MATERIALS AND

METHODS

In this retrospective, multicenter study, 742 female patients (age 49.6 ± 13.2 years) were enrolled from 12 hospitals in China between September 2024 and March 2025. Of these, 557 patients formed the training cohort with five-fold cross-validation and 185 comprised the external validation cohort. Seventy-three radiomic features were extracted from each B-mode image. For each SRUS image, 73 radiomic features and 45 color features were extracted. Clinical variables included patient age, obstetric history, and family history of breast cancer. Feature selection employed Mann-Whitney U testing, Spearman correlation filtering, and least absolute shrinkage and selection operator regression. Support vector machine models were constructed using either single-modality data or multimodal inputs via multi-kernel learning.

RESULTS

The best clinical-radiomic model combining B-mode images, four SRUS maps, and clinical parameters achieved an AUC of 0.921 ± 0.023 internally and 0.872 ± 0.006 externally. In the diagnostically challenging Breast Imaging Reporting and Data System 4A subgroup, it yielded an AUC of 0.876 ± 0.017 and a negative predictive value of 97.5%.

CONCLUSION

The proposed radiomics framework, integrating B-mode US, SRUS, and clinical information, demonstrates robust performance in predicting breast cancer and provides a reference pipeline for SRUS-based microvasculomic analysis.