Deep bayesian approach improves ultrasound recovery in noisy settings
Deep Bayesian REFoCUS
Machine Learning
Summary
Ultrasound imaging relies on interpreting signals that come from sending and receiving sound waves at different points. Sometimes, classic methods to process these signals fail, especially when there isn’t enough information or there is noise. The authors train a deep learning model that understands the patterns in these signals and uses this knowledge to improve image recovery. Their approach not only works better when data is limited or noisy but also estimates how uncertain the recovery is, which classic methods cannot do. Moreover, their method performs well even when moving from simulated data to real patient data without extra adjustments.
What this means in practice
- •For medical imaging engineers: Develop better ultrasound image recovery software that handles limited and noisy acquisition data with improved accuracy and uncertainty awareness.$Commercial implications: Improves ultrasound devices’ image reconstruction reliability, enabling manufacturers to offer advanced diagnostic tools that address challenging clinical scenarios.
- •For industrial non-destructive testing teams: Enhance ultrasound inspection systems to recover signals more accurately in complex setups where classical methods fail due to insufficient data or noise.
Authors
Simon Penninga, Ruud van Sloun
Abstract
In this work we formulate ultrasound multistatic recovery from arbitrary transmit sequences as a Bayesian inference problem. To that end, we train a deep generative prior on multistatic data sets to tackle the rank-deficient regime in which classical linear REFoCUS decoders fail. This appproach, which we term Deep Bayesian REFoCUS, outperforms the linear baselines for all regimes of rank-deficiency and noise levels, and regresses to linear decoding when inversion is exact. The model also expresses uncertainty in the null space of the acquisitions, whereas the linear REFoCUS decoders only provide point estimates. Finally, we analyze the impact of distribution shift between simulation and in-vivo acquisitions, showing remarkable generalization ability without any fine-tuning or adaptation.