Integrating deep learning and radiomics for accurate differentiation of benign and malignant breast...

Mammography (MG) remains the cornerstone imaging modality for early breast cancer screening, playing a pivotal role in lowering disease-specific mortality. Nevertheless, its diagnostic utility is constrained by a substantial false-negative rate, documented between 10% and 30%. This limitation poses significant clinical challenges in evaluating suspicious microcalcifications on surgical excision biopsy. Intraoperative frozen section analysis often yields inconclusive findings for such calcified lesions, potentially necessitating re-operation. Consequently, rigorous assessment of mammographic features in excised specimens is essential for optimizing surgical decision-making, particularly regarding the necessity and timing of intraoperative pathological evaluation.

This investigation involved a retrospective evaluation of 400 intraoperative specimen mammography obtained from individuals confirmed to have breast cancer at the Guangdong Provincial Hospital of Traditional Chinese Medicine. We herein focus exclusively on intraoperative specimen mammography, which is distinct from conventional screening mammography. Upon securing necessary ethical clearance, we assembled a rigorously screened dataset comprising 197 benign and 203 malignant instances. Prior to analysis, images were subjected to preprocessing steps such as grayscale transformation and histogram equalization to optimize them for deep learning ingestion. We subsequently constructed the EfficientNet_radiomics framework by fusing latent representations from the EfficientNet-B0 architecture with manually engineered radiomic descriptors extracted from the imaging data, a strategy designed to augment the system's diagnostic precision.

Compared with conventional machine learning approaches and experienced radiologists, our framework yielded superior diagnostic metrics. Notably, it achieved 78.0% ± 7.0 accuracy in identifying benign calcifications versus 71.4% ± 8.0 for radiologists, and 83.0% ± 6.0 versus 81.8% ± 6.0 for malignant lesions, underscoring its translational value for clinical deployment.

The study holds promise as a real-time decision-support tool during intraoperative specimen evaluation, potentially reducing unnecessary frozen section analyses and re-operation rates by providing more accurate intraoperative differentiation, particularly for benign calcifications.