Enhancing diagnostic precision for BI-RADS 4a breast nodules: A multimodal AI model integrating ultrasound radiomics, hemodynamic signatures, and clinical profiles

Abstract

Background

Breast Imaging Reporting and Data System (BI-RADS) 4a nodules represent a diagnostic dilemma, with a malignancy rate ranging from 2% to 10%. The majority of these nodules prove benign after biopsy, leading to unnecessary invasive procedures, patient anxiety, and healthcare costs. Current clinical practice lacks a reliable, noninvasive tool to accurately distinguish benign from malignant BI-RADS 4a lesions. Emerging evidence suggests that integrating multiparametric ultrasound data with clinical factors through artificial intelligence may improve diagnostic precision, but a validated multimodal model specifically for this ambiguous category is lacking.

Purpose

To develop and validate a multimodal artificial intelligence model that integrates ultrasound radiomics, hemodynamic features, and clinical data for the accurate differentiation of benign and malignant BI-RADS 4a breast nodules, with the ultimate goal of supporting clinical decisionmaking and safely reducing unnecessary biopsies.

Methods

This prospective, multicenter study enrolled 350 pathologically confirmed BI-RADS 4a nodules from 331 patients in one tertiary hospital and two secondary hospitals. Clinical data, grayscale ultrasound images, and hemodynamic parameters were collected for each nodule. Radiomics features were extracted from the ultrasound images. A multimodal fusion model was constructed using the Light Gradient Boosting Machine (LightGBM) algorithm following feature selection via Least Absolute Shrinkage and Selection Operator (LASSO) regression. The model was developed and hyperparameter-tuned on a training set (n = 260) using nested cross-validation and was evaluated on a temporally independent test set (n = 90). Model interpretability was enhanced using SHapley Additive exPlanations (SHAP).

Results

On the independent test set, the multimodal LightGBM model demonstrated excellent diagnostic performance, achieving an area under the receiver operating characteristic curve (AUC) of 0.95 (95% CI: 0.91–0.99), with a sensitivity of 100% (95% CI: 81.5%–100%) and a specificity of 86.1% (95% CI: 75.8%–93.1%). Its performance was significantly superior to baseline classifiers including Logistic Regression, Support Vector Machine, and Random Forest (all p < 0.05). SHAP analysis identified wavelet-transformed texture features, the presence of a penetrating vessel, and BRCA mutation status as the top predictors of malignancy. Applying this model could have potentially avoided 32.4% (24/74) of unnecessary biopsies on benign nodules while maintaining perfect sensitivity.

Conclusion

A multimodal AI model based on LightGBM can accurately differentiate benign and malignant BI-RADS 4a breast nodules. It shows significant potential as a clinical decision-support tool to reduce unnecessary biopsies without compromising sensitivity, enabling more precise management for this challenging patient population.