Parametric Quantitative Ultrasound-Based Deep Learning for Breast Cancer Diagnosis.

Breast cancer is a leading cause of cancer mortality, with improved outcomes as a result of early detection and diagnosis. Breast ultrasound is a valuable clinical modality for detecting breast cancer. However, conventional B-mode imaging is highly operatordependent and suffers from noise and artifacts, reducing its diagnostic accuracy. Traditional computer-aided diagnosis (CAD) systems have shown improved diagnostic accuracy, largely rely on the B-mode image alone and operating as a "closed box," providing little understanding and reasoning behind the prediction. Furthermore, by relying on only the B-mode image, they neglect important acoustic information present in raw ultrasound signals. Going beyond the B-mode image, quantitative ultrasound (QUS) algorithms extract physics-based interpretable parameters from raw radio frequency (RF) ultrasound data. Therefore, we present a physics-based CAD framework to classify breast ultrasound images through the novel combination of parametric QUS images and RF data as inputs to a custom U-Net architecture. We demonstrate that by including all seven features (i.e., mid-band fit (MBF), spectral slope, 0-MHz intercept, and Nakagami-m and Nakagami-Ω, the RF image, and the B-mode image) the model achieved a maximum accuracy of 85% (mean ± standard deviation of 80% ± 4%) and maximum area under the curve (AUC) of 0.90 (mean ± standard deviation of 0.82 ± 0.04), representing an average improvement of 9% compared with grayscale features alone. Furthermore, explainability analyses demonstrate visual differences between benign and malignant heatmaps and showed that Nakagami-m and spectral slope consistently provide the highest predictive significance. These results are promising for the incorporation of physics-based QUS features into deep learning frameworks to improve diagnostic accuracy and enhance explainability, providing a cost-effective, portable solution that can enhance breast cancer diagnosis in resource-limited environments globally.