Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Enhanced Uncertainty Estimation in Ultrasound Image Segmentation with MSU-Net

Rohini Banerjee, Cecilia G. Morales, Artur Dubrawski

TL;DR

This work tackles reliable ultrasound-based vessel segmentation under uncertainty for autonomous needle guidance in austere settings. It introduces MSU-Net, a multistage ensemble that combines bootstrap-generated candidates, decorrelation-based selection, and a final Monte Carlo U-Net combiner to yield accurate segmentation and calibrated uncertainty maps. Compared with a Monte Carlo U-Net baseline, MSU-Net achieves an $18.1\%$ improvement in mean IoU and demonstrates stronger separation between correct and incorrect predictions via Rényi-divergence analysis, enhancing model transparency and trust. The approach has practical implications for safe, non-expert-guided needle insertions in emergency care, with planned validation on live animal and human data.

Abstract

Efficient intravascular access in trauma and critical care significantly impacts patient outcomes. However, the availability of skilled medical personnel in austere environments is often limited. Autonomous robotic ultrasound systems can aid in needle insertion for medication delivery and support non-experts in such tasks. Despite advances in autonomous needle insertion, inaccuracies in vessel segmentation predictions pose risks. Understanding the uncertainty of predictive models in ultrasound imaging is crucial for assessing their reliability. We introduce MSU-Net, a novel multistage approach for training an ensemble of U-Nets to yield accurate ultrasound image segmentation maps. We demonstrate substantial improvements, 18.1% over a single Monte Carlo U-Net, enhancing uncertainty evaluations, model transparency, and trustworthiness. By highlighting areas of model certainty, MSU-Net can guide safe needle insertions, empowering non-experts to accomplish such tasks.

Enhanced Uncertainty Estimation in Ultrasound Image Segmentation with MSU-Net

TL;DR

This work tackles reliable ultrasound-based vessel segmentation under uncertainty for autonomous needle guidance in austere settings. It introduces MSU-Net, a multistage ensemble that combines bootstrap-generated candidates, decorrelation-based selection, and a final Monte Carlo U-Net combiner to yield accurate segmentation and calibrated uncertainty maps. Compared with a Monte Carlo U-Net baseline, MSU-Net achieves an improvement in mean IoU and demonstrates stronger separation between correct and incorrect predictions via Rényi-divergence analysis, enhancing model transparency and trust. The approach has practical implications for safe, non-expert-guided needle insertions in emergency care, with planned validation on live animal and human data.

Abstract

Efficient intravascular access in trauma and critical care significantly impacts patient outcomes. However, the availability of skilled medical personnel in austere environments is often limited. Autonomous robotic ultrasound systems can aid in needle insertion for medication delivery and support non-experts in such tasks. Despite advances in autonomous needle insertion, inaccuracies in vessel segmentation predictions pose risks. Understanding the uncertainty of predictive models in ultrasound imaging is crucial for assessing their reliability. We introduce MSU-Net, a novel multistage approach for training an ensemble of U-Nets to yield accurate ultrasound image segmentation maps. We demonstrate substantial improvements, 18.1% over a single Monte Carlo U-Net, enhancing uncertainty evaluations, model transparency, and trustworthiness. By highlighting areas of model certainty, MSU-Net can guide safe needle insertions, empowering non-experts to accomplish such tasks.
Paper Structure (12 sections, 4 equations, 4 figures, 2 tables, 1 algorithm)

This paper contains 12 sections, 4 equations, 4 figures, 2 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Works
  3. Uncertainty Quantification in Deep Learning
  4. Multistage Neural Network Ensembles
  5. Methods
  6. Monte Carlo U-Net (MCU-Net)
  7. Multistage U-Net (MSU-Net)
  8. Experiments
  9. Results and Discussion
  10. Conclusion
  11. Acknowledgments.
  12. Disclosure of Interests.

Figures (4)

  • Figure 1: Proposed MSU-Net architecture. U-Nets are trained on bootstrap samples and validated on VS1. Decorrelated ensemble members are chosen using VS2.
  • Figure 2: Epistemic uncertainty maps on test data for (a) MCU-Net and (b) MSU-Net. Darker/lighter colors indicate lower/higher uncertainty values. To alleviate class imbalance, we limit evaluations to a region of interest delineated in white.
  • Figure 3: Epistemic uncertainty distributions for correct (blue) and incorrect (orange) predictions for MCU-Net (top) and MSU-Net (bottom). Our approach yields a markedly better differentiation of correct and incorrect predictions.
  • Figure 4: (a) Training (left) and validation (right) curves for MCU-Net and MSU-Net. Early stopping delineated by gray lines. (b) Precision-recall curves.