Radiomic feature-based classification of BI-RADS 4/5 breast lesions on contrast-enhanced mammography.

BACKGROUND AND

OBJECTIVE

Differentiating benign from malignant BI-RADS 4/5 lesions, particularly architectural distortions, remains challenging. We investigated whether radiomics from contrast-enhanced mammography (low-energy, LE, and recombined, RI) improve lesion classification and how performance depends on LE-RI matching, projection, feature selection and classifier choice.

METHODS

We analysed 333 patients with enhancing BI-RADS 4/5 findings (842 focal masses, 184 architectural distortions). Radiomic features were extracted from LE-only and three LE-RI matching strategies. Analyses were performed separately for CC, MLO and combined projections, including projection-specific feature selection. Images were z-score normalised and biopsy markers were removed using patch-based inpainting. Multiple feature selectors and classifiers were evaluated in patient-stratified fivefold cross-validation, and F1, ROC AUC and average precision were compared using linear mixed-effects models.

RESULTS

For focal masses, incorporating RI information improved average precision and ROC AUC by 0.03-0.08 versus LE-only. For architectural distortions, best-match and merged LE-RI strategies yielded larger improvements up to 0.28-0.33 in optimal CC configurations. The impact of RI was strongly projection-dependent, with consistent improvements in the CC view, while effects in MLO were smaller and often not statistically significant. In combined CC+MLO analyses, performing feature selection separately for each projection improved performance, particularly for architectural distortions (up to 0.062 AP, 0.094 F1 and 0.063 ROC AUC). Performance depended on interactions between matching strategy, feature selection method and classifier. GLMM was the most consistent selector for focal masses, whereas merged LASSO-LARS performed best for distortions. The best-performing configurations achieved average precision values of up to approximately 0.86 for architectural distortions and 0.84 for focal masses. YOLOv8-based detection was insufficiently accurate for fully automated radiomics.

CONCLUSIONS

Radiomics from paired LE and RI images improve classification of BI-RADS 4/5 lesions, with performance critically influenced by LE-RI matching and projection-specific feature selection. LE+RI radiomics show promise as a decision-support tool for biopsy-level assessment.