Abstract
Speed-of-Sound (SoS) imaging provides a quantitative biomechanical marker for breast cancer screening, yet reconstructing these maps from pulse-echo ultrasound remains a challenging, ill-posed inverse problem. This paper introduces a framework for trust attribution by integrating uncertainty estimation into Variational Networks (VN). By applying Bayesian techniques—specifically Monte Carlo Dropout (MCD) and Bayesian Variational Inference (BVI)—the researchers can quantify the reliability of reconstructions. A novel relative uncertainty metric is used to automatically select the most trustworthy data frame from multiple clinical acquisitions. Evaluation on a clinical dataset of 21 patients shows that this uncertainty-informed selection boosts diagnostic accuracy (AUC) from 64% to 80%, demonstrating the potential for more reliable AI-assisted breast cancer diagnosis.
Introduction
In clinical ultrasound, Speed-of-Sound (SoS) serves as a proxy for tissue stiffness; malignant tumors typically exhibit higher SoS than benign tissue. However, SoS reconstruction is highly sensitive to noise, patient motion, and acoustic shadows. In practice, a sonographer may capture several scans of a single lesion, but currently lacks an objective method to determine which resulting SoS map is the most accurate. This research addresses the challenge of "trust attribution," enabling a model to identify which of its own outputs are most reliable for clinical decision-making.
Methodology
The authors utilize Variational Networks (VN), a hybrid architecture that unrolls traditional optimization algorithms into neural network layers, combining physics-based priors with data-driven learning. To estimate uncertainty, two Bayesian approaches were implemented:
- Monte Carlo Dropout (MCD): Randomly omitting network components during inference to measure output variance.
- Bayesian Variational Inference (BVI): Treating network filters as probability distributions. The authors introduced an algebraic reformulation to ensure the Kullback–Leibler (KL) loss computation remains efficient for large models.
The core innovation is the Relative Uncertainty Metric. Rather than measuring absolute noise, the system calculates the contrast between uncertainty within a suspected lesion and the surrounding background tissue. This normalized value serves as a surrogate for reconstruction quality, allowing the system to filter out "bad" data frames caused by clinical artifacts.
Results and Discussion
The framework was validated using a pipeline ranging from ray-based and k-Wave simulations to clinical in-vivo data from 21 patients with suspicious (BI-RADS 4) lesions.
Key findings include:
- Diagnostic Improvement: Selecting frames based on BVI uncertainty achieved an AUC of 80% and an F1-score of 80%, a significant jump from the 64% AUC achieved by uninformed baseline selection.
- Relative vs. Absolute Metrics: The study found that while absolute uncertainty varies across patients, the contrast in uncertainty between the lesion and background is a robust indicator of reconstruction quality.
- Regularization Benefits: The addition of Bayesian constraints did not degrade image quality; in some instances, it acted as a regularizer that improved the final SoS maps.
Conclusion
By treating uncertainty as a selection tool rather than just an error metric, this work bridges the gap between deep learning models and clinical trust. While limited by a small clinical sample size and the current need for manual lesion masking, the results suggest that uncertainty-aware Variational Networks can significantly improve the reliability of quantitative ultrasound imaging in breast cancer diagnostics.
Summary updated