An interpretable ultrasound model combining quantitative image features with clinical information improved the classification of benign and malignant breast lesions and may help refine biopsy decisions. Published in Insights into Imaging, the multicentre study included patients from eight hospitals in China, with retrospective development and validation alongside a prospective test cohort. The combined approach performed better than models based only on image features or clinical ultrasound findings. It also remained consistent across age and lesion-size groups. Reclassification of selected Breast Imaging Reporting and Data System categories reduced unnecessary biopsy recommendations without a significant loss of sensitivity, although broader prospective assessment is still required.
Must Read: Axillary US Helps Assess Nodal Burden in Breast Cancer
Combining Imaging and Clinical Information
The analysis included more than 3,000 adults with ultrasound-detected breast lesions confirmed by pathology. Patient data were divided into training, internal testing, external testing and prospective testing groups. The external cohort came from several hospitals, while the prospective cohort was drawn from one centre. One lesion was assessed per patient, with the largest lesion selected when several were present. Ultrasound examinations were performed within two weeks before biopsy or surgery, and the latest eligible scan was analysed.
Radiologists manually marked lesion boundaries on stored ultrasound images. Automated processing then extracted a large set of quantitative features related to shape, intensity and texture. Reproducibility and feature-selection steps reduced these to a smaller group of representative measures. The selected image features were also tested for consistency between repeated segmentations.
Five machine learning methods were compared, including logistic regression and random forest. Random forest performed perfectly in the training data but declined markedly in the test groups, indicating limited generalisability. Logistic regression showed more stable performance and was selected for the radiomics component. A separate clinical model used age, lesion size, orientation, margin and shape, which remained associated with malignancy. The final model combined the radiomics score with these clinical and ultrasound features. This integration brought together reader-assessed characteristics and quantitative information derived from the images, while avoiding use of the final BI-RADS category as an input.
Higher Performance Across Test Groups
The combined model consistently outperformed the radiomics-only and clinical-only approaches. Its area under the curve remained around 0.90 or higher across the training, internal, external and prospective datasets. The clinical model performed well but remained below the combined approach, while the radiomics model showed lower but stable discrimination. Performance also remained balanced in subgroups based on age and lesion size.
Across these subgroup analyses, values ranged from 0.87 to 0.94. Results were consistent among younger patients and those with smaller lesions, with no major loss of discrimination across the tested groups. Calibration was satisfactory, and predicted risks separated benign from malignant lesions across all datasets. Decision curve analysis showed greater net benefit for the combined model than for either component alone across a broad range of decision thresholds.
Additional reclassification measures also indicated that the radiomics score added useful information beyond conventional clinical and ultrasound features. Interpretability was addressed through feature-contribution analysis. This ranked the influence of the radiomics score, age, margin, shape, lesion size and orientation. The radiomics score had the strongest contribution, followed by several clinical and ultrasound variables. Individual case displays showed which features shifted a prediction towards benignity or malignancy. The analysis therefore provided both overall and case-level explanations of how the included variables contributed to model predictions.
Potential to Refine Biopsy Recommendations
The combined model was compared with assessments made by experienced radiologists using different BI-RADS thresholds. At the lower threshold, radiologist assessment produced very high sensitivity but very low specificity. The model improved overall accuracy and specificity but showed lower sensitivity. At intermediate and higher thresholds, differences in overall accuracy were smaller, although the balance between sensitivity and specificity varied.
The main clinical utility analysis focused on BI-RADS categories 3 and 4a. Lesions classified as low risk by the model could move from 4a to 3, while high-risk category 3 lesions could move to 4a. After this adjustment, unnecessary biopsy rates fell by about 10% in both the external and prospective test groups. Sensitivity did not fall significantly, and the proportion of malignant lesions within category 4a increased.
These findings require cautious interpretation. All included lesions had pathological confirmation, creating a selected cohort with a relatively high prevalence of malignancy. The population therefore did not fully reflect routine diagnostic ultrasound practice, particularly lower-prevalence settings. Ultrasound equipment and imaging protocols also varied between centres. Lesion boundaries were marked manually, and the combined model included reader-derived descriptors such as margin, shape and orientation. The comparison with radiologist assessment therefore did not represent a completely automated human-versus-machine evaluation. Further prospective clinical utility testing is needed before routine implementation.
Combining ultrasound radiomics with age and conventional lesion characteristics produced stronger and more consistent classification than either approach alone. The model remained stable across test groups, offered visible explanations for individual predictions and showed potential to reduce unnecessary biopsy recommendations through targeted adjustment of equivocal BI-RADS categories. However, the selected pathology-confirmed cohort, variation in imaging acquisition and reliance on manual segmentation limit immediate use in routine care. Broader prospective evaluation is needed to establish whether the diagnostic and biopsy-related benefits can be reproduced in more representative populations and across different clinical settings.
Source: Insights into Imaging
Image Credit: iStock
References:
Zhang D, Lu WW, Qin XC et al. (2026) Development and validation of an interpretable ultrasound radiomics model for benign and malignant classification of breast lesions: a multicenter large-sample study. Insights Imaging; 17, 174.
