Can Multimodal Ultrasound Refine BI-RADS 4a Decisions?

A difficult middle category

BI-RADS 4a breast nodules are considered suspicious enough to warrant assessment, yet many prove benign. A non-invasive model that helps distinguish risk could support decisions, but it must be evaluated carefully: missing a cancer and recommending an unnecessary biopsy have different consequences.

A prospective, multicentre study in the Journal of Applied Clinical Medical Physics combined ultrasound radiomics with hemodynamic and clinical features. The cohort included 350 pathologically confirmed nodules from 331 patients at three hospitals. The model was trained on 260 nodules and evaluated on a temporally independent set of 90.

Performance and model interpretation

On the test set, the LightGBM model achieved an area under the ROC curve of 0.95. Reported sensitivity was 100% (95% CI 81.5%–100%) and specificity was 86.1% (95% CI 75.8%–93.1%). SHAP analysis highlighted texture features, a penetrating vessel, and BRCA mutation status among the model’s leading predictors.

The authors estimated that, in this cohort, the model could have avoided 24 of 74 biopsies of benign nodules while retaining the observed sensitivity. That is a retrospective application of the model’s predictions to the study data, not evidence that biopsies can safely be skipped in routine care.

The uncertainty behind a strong score

The independent test set was small, and the confidence interval around sensitivity is wide. The study’s participants had pathology-confirmed outcomes, and prospective use in other populations and imaging settings remains to be established. SHAP explanations describe which inputs influenced the model; they do not prove that those inputs are causal or that the model’s reasoning is clinically correct.

The findings support further validation, not a change to screening or biopsy guidance. Larger external studies should assess calibration, subgroup performance, and the consequences of using the model in real decisions.

Source and credit

Independent summary of “Enhancing diagnostic precision for BI-RADS 4a breast nodules: A multimodal AI model integrating ultrasound radiomics, hemodynamic signatures, and clinical profiles,” Journal of Applied Clinical Medical Physics, published 1 October 2026. Original paper, DOI: 10.1002/acm2.70800. Source article license: CC BY 4.0. No source figures or tables are reproduced here.

Questions about this article? Contact the author. Research discussion is welcome; please do not submit personal medical information.