Dual-view and dual-modality fusion in contrast-enhanced spectral mammography for breast-side classification: a patient-grouped evaluation.

AIM: To develop and evaluate a patient-grouped breast-side classifier for contrast-enhanced spectral mammography that integrates information across both standard projections and both routinely interpreted image types, and to report calibrated probability outputs with uncertainty estimates.

MATERIALS AND

METHODS

The public categorised digital database for contrast-enhanced spectral mammography (CDD-CESM) dataset (326 patients; 2006 images) was analysed at the breast-side level (566 sides) using a binary endpoint (normal vs abnormal; benign and malignant grouped as abnormal). Each breast side was represented by up to four images: two projections (craniocaudal and mediolateral oblique) and two image types (low-energy and subtracted). A transformer-based fusion network with explicit handling of missing inputs was evaluated using five-fold patient-grouped cross-validation (277 patients) and a patient-disjoint holdout set (49 patients; 85 sides). Predictive probabilities were derived using Monte Carlo dropout (MC-dropout) with 10 sample passes and calibrated within each fold using validation-fitted logistic calibration; decision thresholds were selected on validation data and applied unchanged to test folds and the holdout set.

RESULTS

In cross-validation, the model achieved receiver operating characteristic-area under the curve (ROC-AUC) 0.993 ± 0.008 and area under the precision-recall curve (PR-AUC) 0.996 ± 0.004, with sensitivity 0.971 ± 0.034 and specificity 0.950 ± 0.029 at the validation-selected threshold (F1-score 0.973 ± 0.022; accuracy 0.964 ± 0.029). Calibration metrics were Brier score 0.025 ± 0.017, expected calibration error 0.024 ± 0.009, and negative log-likelihood 0.144 ± 0.092. On the holdout set, ROC-AUC was 0.985 with sensitivity of 0.984 and specificity of 0.792 (F1-score 0.953; accuracy 0.930).

CONCLUSION

Patient-grouped evaluation showed that dual-view, dual-modality fusion of contrast-enhanced spectral mammography can provide strong breast-side discrimination with calibrated probability outputs, while remaining robust to incomplete inputs.