Evidential DualU-Net: Single-Pass Uncertainty for Cell Instance Segmentation

David Anglada-Rotger, Ferran Marques, Montse Pardàs

Keywords: Cell Instance Segmentation, Evidential Deep Learning, Uncertainty Estimation, Multitask Learning, Nuclei Segmentation

TL;DR: Evidential DualU-Net adds single-pass uncertainty estimation to DualU-Net using Dirichlet outputs and centroid cues. It provides calibrated, instance-level uncertainties, matches ensembles at lower cost, and generalizes well across datasets.

Abstract: Accurate and trustworthy cell instance segmentation requires models that not only detect and classify nuclei but also communicate how much evidence supports each prediction. DualU-Net is a fast and effective two-head multi-task architecture for this problem, but—like most deterministic models—it provides no principled uncertainty estimates. We introduce Evidential DualU-Net, the first evidential framework for multi-task cell instance segmentation.Its segmentation head predicts Dirichlet concentration parameters, enabling single-pass, closed-form aleatoric, epistemic, and vacuity uncertainties at the pixel level, with instance-level quantities obtained via size-invariant pooling of pixel evidence. The centroid decoder is complemented with two lightweight geometric uncertainty cues that quantify localisation reliability without auxiliary models or sampling. Together, these evidential and geometric measures expose complementary failure modes and allow principled filtering of low-confidence nuclei. Across multi-tissue and multi-stain datasets, Evidential DualU-Net matches or surpasses deep ensembles and MC Dropout in error separation at a fraction of the cost, maintains or improves calibration over deterministic baselines, and generalises across datasets without retuning. This work provides an interpretable and computationally practical uncertainty formulation for digital pathology. Code and weights are available at: https://github.com/davidanglada/Evidential-DualU-Net.

Primary Subject Area: Uncertainty Estimation

Secondary Subject Area: Segmentation

Registration Requirement: Yes

Reproducibility: https://github.com/davidanglada/Evidential-DualU-Net

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Midl Latex Submission Checklist: Ensure no LaTeX errors during compilation., Replace NNN with your OpenReview submission ID., Includes \documentclass{midl}, \jmlryear{2026}, \jmlrworkshop, \jmlrvolume, \editors, and correct \bibliography command., Did not override options of the hyperref package., Did not use the times package., Use the correct spelling and format, avoid Unicode characters, and use LaTeX equivalents instead., Any math in the title and abstract must be enclosed within $...$., Did not override the bibliography style defined in midl.cls and did not use \begin{thebibliography} directly to insert references., Avoid using \scalebox; use \resizebox when needed., Included all necessary figures and removed *unused* files in the zip archive., Removed special formatting, visual annotations, and highlights used during rebuttal., All special characters in the paper and .bib file use LaTeX commands (e.g., \'e for é)., No separate supplementary PDF uploads., Acknowledgements, references, and appendix must start after the main content.

Submission Number: 285