Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Uncertainty-Error correlations in Evidential Deep Learning models for biomedical segmentation

Hai Siong Tan, Kuancheng Wang, Rafe Mcbeth

TL;DR

The paper addresses uncertainty quantification in biomedical segmentation by applying Evidential Deep Learning with Dirichlet priors to two MRI datasets (cardiac and prostate). It compares EDl against Shannon-entropy baselines, MC Dropout, and Deep Ensembles, showing that EDL yields stronger uncertainty-error correlations while preserving Dice accuracy. It also demonstrates the utility of EDL uncertainty heatmaps for identifying potential errors and shows that EDL-based uncertainty sampling improves active-learning performance without compromising final segmentation quality. Overall, the study supports deploying EDL as a robust, uncertainty-aware approach for error-sensitive clinical segmentation tasks and for guiding expert intervention and data acquisition.

Abstract

In this work, we examine the effectiveness of an uncertainty quantification framework known as Evidential Deep Learning applied in the context of biomedical image segmentation. This class of models involves assigning Dirichlet distributions as priors for segmentation labels, and enables a few distinct definitions of model uncertainties. Using the cardiac and prostate MRI images available in the Medical Segmentation Decathlon for validation, we found that Evidential Deep Learning models with U-Net backbones generally yielded superior correlations between prediction errors and uncertainties relative to the conventional baseline equipped with Shannon entropy measure, Monte-Carlo Dropout and Deep Ensemble methods. We also examined these models' effectiveness in active learning, finding that relative to the standard Shannon entropy-based sampling, they yielded higher point-biserial uncertainty-error correlations while attaining similar performances in Dice-Sorensen coefficients. These superior features of EDL models render them well-suited for segmentation tasks that warrant a critical sensitivity in detecting large model errors.

Uncertainty-Error correlations in Evidential Deep Learning models for biomedical segmentation

TL;DR

The paper addresses uncertainty quantification in biomedical segmentation by applying Evidential Deep Learning with Dirichlet priors to two MRI datasets (cardiac and prostate). It compares EDl against Shannon-entropy baselines, MC Dropout, and Deep Ensembles, showing that EDL yields stronger uncertainty-error correlations while preserving Dice accuracy. It also demonstrates the utility of EDL uncertainty heatmaps for identifying potential errors and shows that EDL-based uncertainty sampling improves active-learning performance without compromising final segmentation quality. Overall, the study supports deploying EDL as a robust, uncertainty-aware approach for error-sensitive clinical segmentation tasks and for guiding expert intervention and data acquisition.

Abstract

In this work, we examine the effectiveness of an uncertainty quantification framework known as Evidential Deep Learning applied in the context of biomedical image segmentation. This class of models involves assigning Dirichlet distributions as priors for segmentation labels, and enables a few distinct definitions of model uncertainties. Using the cardiac and prostate MRI images available in the Medical Segmentation Decathlon for validation, we found that Evidential Deep Learning models with U-Net backbones generally yielded superior correlations between prediction errors and uncertainties relative to the conventional baseline equipped with Shannon entropy measure, Monte-Carlo Dropout and Deep Ensemble methods. We also examined these models' effectiveness in active learning, finding that relative to the standard Shannon entropy-based sampling, they yielded higher point-biserial uncertainty-error correlations while attaining similar performances in Dice-Sorensen coefficients. These superior features of EDL models render them well-suited for segmentation tasks that warrant a critical sensitivity in detecting large model errors.

Paper Structure

This paper contains 12 sections, 19 equations, 4 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Preliminaries
  3. Methodology
  4. On datasets, model structure and loss function
  5. On active learning
  6. On measures of uncertainty-error correlations
  7. Results
  8. EDL showed the strongest uncertainty-error correlation
  9. On uncertainty heatmaps
  10. On active learning
  11. Discussion
  12. Conclusion

Figures (4)

  • Figure 1: Empirical cumulative distribution functions (eCDF) of uncertainties in various models trained for prostate segmentation. Highly similar trends were also observed for eCDFs pertaining to the cardiac (left atrium) segmentation.
  • Figure 2: Aleatoric uncertainty heatmaps for segmentation of the left atrium in sagittal T1-weighted cardiac MRI images. Red curves pertain to model-predicted contours overlaid on ground truth masks in yellow.
  • Figure 3: Epistemic uncertainty heatmaps for segmentation of the central and peripheral zones of the prostate in axial T2-weighted MRI images. Note how elevated redder regions of each heatmap appear to mirror deviations between ground truth masks in yellow and their model-predicted contours in red.
  • Figure 4: Evolution of Dice coefficient during active training.