Research map / 3D vision / Depth and stereo

Evidential Uncertainty-Guided Mitochondria Segmentation for 3D EM Images

AAAI 2024 · 3D vision › Depth and stereo · approach: Feed-forward 3D reconstruction · 9 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes EUMS-3D, a 3D mitochondria instance segmentation network that adds evidential deep learning uncertainty estimation and an uncertainty rectification module with multi-scale feature aggregation, built on 3D UX-Net.

Problem

Mitochondria segmentation in 3D EM images produces overconfident errors in high-uncertainty regions (ambiguous boundaries, false positive segments), limiting reliability for 3D reconstruction.

Approach

Use evidential deep learning (Dirichlet-based subjective logic) to estimate voxel-wise aleatoric uncertainty on semantic and boundary predictions, then rectify predictions with an attention-based module that fuses uncertainty maps and multi-scale aggregated features, rather than only using uncertainty for fusion or pseudo-labeling.

Key results

• EUMS-3D (3D UX-Net backbone) reaches 0.845 Dice / 0.901 mAP on MitoEM-H and 0.890 Dice / 0.928 mAP on MitoEM-R, outperforming the 3D UX-Net baseline (0.816/0.845 and 0.859/0.901).
• On Kasthuri++ and Lucchi++, EUMS-3D achieves 0.972 Dice / 0.931 mAP and 0.937 Dice / 0.953 mAP respectively.
• Evidential uncertainty estimation outperforms dropout and ensemble alternatives on MitoEM-H/R (e.g., mAP 0.901 vs 0.873 dropout and 0.880 ensemble on MitoEM-H).

Measured effects reported in the paper

Surprising findings

• Removing the uncertainty rectification module drops MitoEM-R mAP by about 2.2 points, with the gain attributed to uncertainty-guided correction of false positives in unannotated mitochondria-like structures.
• Replacing 3D convolutions with dilated convolutions slightly reduces performance, suggesting a broader receptive field adds noise.

Limitations

• Uncertainty only models aleatoric uncertainty; epistemic uncertainty is not addressed.
• Gains are reported on a single backbone for the full method; the threshold tau is tuned on MitoEM-R.

Experimental setup

Datasets: MitoEM-R MitoEM-H Kasthuri++ Lucchi++

Benchmarks: MitoEM-H MitoEM-R Kasthuri++ Lucchi++

Metrics: Dice mAP NPV TPR UA (uncertainty accuracy)

Baselines: Res-Unet3D HIVE-Net Zhili Peng and Yuan Xiao nnU-Net DSTUnet TransBTS UNETR nnFormer SwinUNETR 3D UX-Net

Builds on: 3D UX-Net evidential deep learning Mask2Former-style mask decoder soft Dice loss

Closest papers by meaning

• Alleviate and Mining: Rethinking Unsupervised Domain Adaptation for Mitochondria Segmentation from Pseudo-Label Perspective (AAAI 2025 · similarity 0.89)
• Electron Microscopy Images as Set of Fragments for Mitochondrial Segmentation (AAAI 2024 · similarity 0.87)
• $\epsilon$-Seg: Sparsely Supervised Semantic Segmentation of Microscopy Data (NeurIPS 2025 · similarity 0.83)
• DyCON: Dynamic Uncertainty-aware Consistency and Contrastive Learning for Semi-supervised Medical Image Segmentation (CVPR 2025 · similarity 0.83)
• Arctique: An artificial histopathological dataset unifying realism and controllability for uncertainty quantification (NeurIPS 2024 · similarity 0.82)
• Multi-view Evidential Learning-based Medical Image Segmentation (AAAI 2025 · similarity 0.81)