Uncertainty Estimation in Speed-of-Sound Reconstruction Using Variational Networks for Breast Cancer Imaging

Overview

This study introduces a novel uncertainty estimation framework for speed-of-sound (SoS) reconstruction in ultrasound imaging using variational networks (VN). The approach enables trust attribution to select optimal acquisition frames, improving the reliability of SoS imaging for breast cancer diagnosis, demonstrated for the first time in vivo.

Background

Breast cancer remains a leading cause of mortality among women, with early detection critical for reducing death rates. Conventional imaging modalities like mammography and MRI have limitations including ionizing radiation exposure and reduced sensitivity in dense breast tissue. Ultrasound offers a non-ionizing, real-time alternative, but conventional B-mode imaging lacks sufficient contrast to differentiate malignant from benign lesions. Speed-of-sound imaging, which quantifies tissue biomechanical properties, shows promise for lesion characterization but requires advanced reconstruction techniques due to the ill-posed nature of the inverse problem. Variational networks provide a model-based deep learning framework that can generalize from simulated to in vivo data, enhancing SoS reconstruction.

Data Highlights

The study utilizes multiple beamformed ultrasound frames with different transmit/receive parameters to estimate local speckle shifts sensitive to SoS variations. The inverse problem is formulated as minimizing a data fidelity term plus a regularization term, solved via variational networks trained with Bayesian variational inference and Monte Carlo Dropout. Uncertainty estimates are derived from posterior samples, with the mean representing the SoS image and the standard deviation quantifying uncertainty. A novel relative uncertainty metric is introduced to improve trust attribution in frame selection. The method was validated on clinical breast lesion data, demonstrating improved acquisition selection and diagnostic potential.

Key Findings

• First application of uncertainty estimation to speed-of-sound reconstruction in ultrasound imaging using variational networks.
• Introduction of a relative (normalized) uncertainty metric that outperforms absolute uncertainty in trust attribution for regression tasks.
• Demonstration that uncertainty estimates can guide selection of optimal acquisition frames, leveraging ultrasound's rapid and interactive imaging capability.
• Successful in vivo application of variational networks for SoS reconstruction in breast cancer diagnosis, differentiating malignant ductal carcinoma from benign fibroadenoma.
• Development of an algebraic reformulation enabling efficient uncertainty computation for large models on standard clinical hardware.

Clinical Implications

Incorporating uncertainty estimation into SoS reconstruction enhances the reliability of ultrasound-based breast lesion characterization, potentially reducing false diagnoses. The ability to select optimal acquisition frames based on trust attribution can improve diagnostic confidence and streamline clinical workflows. This approach supports the broader adoption of SoS imaging as a non-ionizing, real-time adjunct to conventional breast cancer screening methods.

Conclusion

This work advances speed-of-sound ultrasound imaging by integrating uncertainty estimation via variational networks, enabling trustworthy and clinically feasible breast cancer diagnosis. The proposed framework paves the way for more reliable, interactive, and efficient ultrasound imaging protocols in oncology.

References

1. Breast cancer mortality and lesion types [1,2]
2. Limitations of mammography and MRI [3]
3. Speed-of-sound imaging and reconstruction methods [4-13]
4. Variational networks and model-based deep learning [16-20]
5. Uncertainty estimation frameworks in medical imaging [21-24]