Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Deep Sylvester Posterior Inference for Adaptive Compressed Sensing in Ultrasound Imaging

Simon W. Penninga, Hans van Gorp, Ruud J. G. van Sloun

TL;DR

Ultrasound imaging faces a trade-off between frame rate, depth, and image quality due to limited transmit events. The paper introduces an active subsampling framework that combines a deep generative latent-variable model with a Sylvester normalizing-flow posterior encoder to approximate the Bayesian posterior $p(z_t|y_t)$ under partial observations, and a mutual-information-based policy to select the next scan-lines $A_{t+1}$. Key contributions include fast posterior inference, two information-gain based sampling strategies (covariance and trace variants), and a thorough evaluation on the EchoNet dataset showing up to a $15\%$ improvement in reconstruction under aggressive subsampling with real-time performance around $0.015$ s per frame ($\approx66$ Hz). The approach enables real-time adaptive sensing in ultrasound, offering substantial gains in frame rate and energy efficiency, with potential extensions to memory-enabled architectures and 3D imaging.

Abstract

Ultrasound images are commonly formed by sequential acquisition of beam-steered scan-lines. Minimizing the number of required scan-lines can significantly enhance frame rate, field of view, energy efficiency, and data transfer speeds. Existing approaches typically use static subsampling schemes in combination with sparsity-based or, more recently, deep-learning-based recovery. In this work, we introduce an adaptive subsampling method that maximizes intrinsic information gain in-situ, employing a Sylvester Normalizing Flow encoder to infer an approximate Bayesian posterior under partial observation in real-time. Using the Bayesian posterior and a deep generative model for future observations, we determine the subsampling scheme that maximizes the mutual information between the subsampled observations, and the next frame of the video. We evaluate our approach using the EchoNet cardiac ultrasound video dataset and demonstrate that our active sampling method outperforms competitive baselines, including uniform and variable-density random sampling, as well as equidistantly spaced scan-lines, improving mean absolute reconstruction error by 15%. Moreover, posterior inference and the sampling scheme generation are performed in just 0.015 seconds (66Hz), making it fast enough for real-time 2D ultrasound imaging applications.

Deep Sylvester Posterior Inference for Adaptive Compressed Sensing in Ultrasound Imaging

TL;DR

Ultrasound imaging faces a trade-off between frame rate, depth, and image quality due to limited transmit events. The paper introduces an active subsampling framework that combines a deep generative latent-variable model with a Sylvester normalizing-flow posterior encoder to approximate the Bayesian posterior under partial observations, and a mutual-information-based policy to select the next scan-lines . Key contributions include fast posterior inference, two information-gain based sampling strategies (covariance and trace variants), and a thorough evaluation on the EchoNet dataset showing up to a improvement in reconstruction under aggressive subsampling with real-time performance around s per frame ( Hz). The approach enables real-time adaptive sensing in ultrasound, offering substantial gains in frame rate and energy efficiency, with potential extensions to memory-enabled architectures and 3D imaging.

Abstract

Ultrasound images are commonly formed by sequential acquisition of beam-steered scan-lines. Minimizing the number of required scan-lines can significantly enhance frame rate, field of view, energy efficiency, and data transfer speeds. Existing approaches typically use static subsampling schemes in combination with sparsity-based or, more recently, deep-learning-based recovery. In this work, we introduce an adaptive subsampling method that maximizes intrinsic information gain in-situ, employing a Sylvester Normalizing Flow encoder to infer an approximate Bayesian posterior under partial observation in real-time. Using the Bayesian posterior and a deep generative model for future observations, we determine the subsampling scheme that maximizes the mutual information between the subsampled observations, and the next frame of the video. We evaluate our approach using the EchoNet cardiac ultrasound video dataset and demonstrate that our active sampling method outperforms competitive baselines, including uniform and variable-density random sampling, as well as equidistantly spaced scan-lines, improving mean absolute reconstruction error by 15%. Moreover, posterior inference and the sampling scheme generation are performed in just 0.015 seconds (66Hz), making it fast enough for real-time 2D ultrasound imaging applications.
Paper Structure (9 sections, 7 equations, 2 figures, 1 table, 1 algorithm)

This paper contains 9 sections, 7 equations, 2 figures, 1 table, 1 algorithm.

Table of Contents

  1. Introduction
  2. Methods
  3. Problem setup
  4. Deep Sylvester Posterior Inference
  5. Sampling Policy
  6. Experiments & Results
  7. Experimental Setup
  8. Results
  9. Discussion and conclusion

Figures (2)

  • Figure 1: Schematic overview of the active sampling loop of a single video frame. Partial observations of the full frame are used to estimate the latent posterior distribution of the next frame of the video. Samples from this posterior distribution are used to estimate mutual information between the state and the observation, which in turn determines the next subsampling mask and results in new observations.
  • Figure 2: Reconstruction results for three consecutive frames $t_{11}, t_{12}$ and $t_{13}$ of an ultrasound video that has median performance for trace-sampling (L1-Loss = 0.070). The final column shows the representation limit that is given by the deep generative model (L1-Loss = 0.055). The smaller blue cones show the absolute difference between the posterior mean and the ground truth.