Multifeature Ultrasound-Based Classification for Breast Lesions: A Comparative Study of PONS Image Enhancement...

To overcome critical limitations of B-mode ultrasound in artificial intelligence diagnostics-including poor image quality and operator variability-by developing a multifeature framework that combines raw B-mode scans with 2 optimized representations (enhanced ultrasound and quality-improved ultrasound) for robust breast cancer classification.

We conducted a retrospective study of 62,912 breast ultrasound scans (100%) from 688 patients (100%) at the Mayo Clinic (from December 01, 1989 to March 30, 2024). The study compared 3 deep learning architectures-graph convolutional networks (GCNs), masked autoencoders (MAEs), and multi-scale convolutional neural network (MSCNN)-using either standard B-mode inputs alone or combined with our enhanced features. Performance was evaluated through 3-fold cross-validation at the patient-level, with primary metrics including accuracy, area under the curve, F1-score, sensitivity, and specificity.

The multifeature approach reported substantial improvements across all metrics. For GCNs, multifeature integration increased accuracy from 0.508 to 0.845 and F1-score from 0.067 to 0.835. Sensitivity improved dramatically from 5.6% to 91.7%, while specificity showed a modest decrease from 85.7% to 79.0%. The MAE models showed different but complementary strengths, with multifeature integration improving accuracy from 0.775 to 0.873, F1-score from 0.785 to 0.822, and achieving perfect specificity (100%) while maintaining clinically acceptable sensitivity (71.1%). The MSCNN, included as a baseline convolutional architecture, showed minimal improvement with multifeature integration, with accuracy increasing slightly from 0.571 to 0.585 and specificity from 0.667 to 0.819. These results highlight the superior capability of GCNs and MAEs to effectively leverage multifeature information in breast ultrasound analysis compared with conventional MSCNN.

PONS-enhanced multifeature ultrasound significantly improves breast cancer detection accuracy versus B-mode alone, offering complementary clinical solutions: GCNs for high sensitivity screening (91.7%) and MAEs for high-specificity diagnosis (100%). Results demonstrate clinical potential across diverse populations, with future work exploring enhanced fusion strategies.