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

Area of Science:

• Medical Imaging
• Artificial Intelligence in Diagnostics
• Oncology

Background:

• B-mode ultrasound limitations in AI diagnostics include poor image quality and operator variability.
• Developing robust AI diagnostic tools requires overcoming these limitations for accurate breast cancer classification.

Purpose of the Study:

• To develop and evaluate a multifeature framework combining raw B-mode ultrasound with enhanced and quality-improved representations.
• To improve the robustness and accuracy of AI-driven breast cancer classification.

Main Methods:

• A retrospective study analyzed 62,912 breast ultrasound scans from 688 patients.
• Compared three deep learning architectures (GCNs, MAEs, MSCNN) using B-mode inputs alone versus combined with enhanced features.
• Performance evaluated using 3-fold cross-validation with metrics including accuracy, AUC, F1-score, sensitivity, and specificity.

Main Results:

• The multifeature approach significantly improved performance across GCNs and MAEs.
• GCNs with multifeature integration showed accuracy from 0.508 to 0.845 and sensitivity from 5.6% to 91.7%.
• MAEs achieved accuracy from 0.775 to 0.873 and perfect specificity (100%) with multifeature integration.

Conclusions:

• PONS-enhanced multifeature ultrasound significantly improves breast cancer detection accuracy over B-mode alone.
• GCNs and MAEs effectively leverage multifeature information, offering complementary strengths for screening and diagnosis.
• The framework shows clinical potential for diverse populations, with future work on enhanced fusion strategies.