Research map / Detection and segmentation / Open-set and few-shot detection
Arctique: An artificial histopathological dataset unifying realism and controllability for uncertainty quantification
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.
Introduces Arctique, a Blender-based procedurally generated histopathology-like dataset with controllable image and label noise, and uses it to benchmark MSR, TTA, MC dropout and deep ensembles for uncertainty quantification in segmentation.
Problem
Existing uncertainty quantification (UQ) benchmarks for image segmentation are either overly simple toy datasets or complex real datasets where true uncertainty cannot be controlled or disentangled.
Approach
A procedurally generated 3D colon histopathology scene pipeline in Blender with independent sliders for image-level manipulation (nuclei staining intensity, blood-cell abundance) and label-level noise (class switching, shape corruption, missing labels), enabling controlled study of aleatoric and epistemic uncertainty.
Key results
- Manipulating label noise increases predictive uncertainty across all four UQ methods, with the majority of uncertainty attributed to the aleatoric component for both FG-BG and semantic segmentation.
- Models trained only on Arctique achieve zero-shot F1 and Hausdorff results on real H&E data (Lizard), with per-class correlations reported for epithelial cells and lymphocytes.
- Blood-stain image manipulation markedly reduces accuracy (misclassifying blood cells as eosinophils) while uncertainty measures barely change, showing UQ methods can fail to flag errors.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Increase label noise by randomly switching class labels | Segmentation models on Arctique Sem-Seg · Arctique Sem-Seg | accuracy | Hurt |
| Increase label noise by randomly switching class labels | UQ methods on Arctique Sem-Seg · Arctique Sem-Seg | predictive uncertainty | Helped |
| Increase label noise by modifying or removing cell labels | Segmentation models on Arctique FG-BG-Seg · Arctique FG-BG-Seg | accuracy | Hurt |
| Increase label noise by modifying or removing cell labels | UQ methods on Arctique FG-BG-Seg · Arctique FG-BG-Seg | predictive uncertainty | Helped |
| Increase label noise | Deep Ensembles, Test Time Augmentation, and Monte-Carlo Dropout on Arctique · Arctique label-noise scenarios | aleatoric uncertainty | Helped |
| Reduce nuclei staining intensity | Segmentation models on Arctique FG-BG-Seg · Arctique Nuclei-Intensity | accuracy | Hurt |
| Reduce nuclei staining intensity | UQ methods on Arctique FG-BG-Seg · Arctique Nuclei-Intensity | uncertainty | Helped |
| Increase blood-cell abundance and red staining | Segmentation models on Arctique Sem-Seg · Arctique Blood-Stain | accuracy | Hurt |
| Increase blood-cell abundance and red staining | UQ methods on Arctique Sem-Seg · Arctique Blood-Stain | uncertainty | No clear effect |
| Increase blood-cell abundance and red staining | Deep Ensembles on Arctique Sem-Seg · Arctique Blood-Stain | uncertainty | Hurt |
Surprising findings
- For the Blood-Stain manipulation, error rates rise sharply while uncertainty stays low or even decreases for deep ensembles.
- Missing labels in densely packed regions are interpolated by the model and not reflected in uncertainty maps.
Limitations
- Limited to a single tissue type (colon H&E) and a restricted set of five cell types.
- Only four UQ methods and a UNet backbone are benchmarked, as a proof of concept.
- Realism is assessed indirectly through zero-shot segmentation of real data rather than a full distribution comparison.
Experimental setup
Datasets: Arctique Lizard
Benchmarks: Arctique
Metrics: accuracy F1 score Hausdorff distance predictive uncertainty epistemic uncertainty aleatoric uncertainty
Baselines: Maximum Softmax Response Test Time Augmentation Monte-Carlo Dropout Deep Ensembles
Builds on: Lizard dataset HoVer-NeXt UNet Monte-Carlo Dropout Deep Ensembles
Closest papers by meaning
- STREAM: Stochastic Riemannian Flow Matching with Anisotropic Decoder for Digital Histopathology Image Generation (NeurIPS 2026 · similarity 0.86)
- PathVQ: Reforming Computational Pathology Foundation Model for Whole Slide Image Analysis via Vector Quantization (NeurIPS 2025 · similarity 0.85)
- DyCON: Dynamic Uncertainty-aware Consistency and Contrastive Learning for Semi-supervised Medical Image Segmentation (CVPR 2025 · similarity 0.84)
- Semantic and Visual Crop-Guided Diffusion Models for Heterogeneous Tissue Synthesis in Histopathology (NeurIPS 2025 · similarity 0.84)
- PathView-Bench: Can Multimodal Large Language Models Achieve Fine-grained Multiscale Understanding of Pathology Images? (NeurIPS 2026 · similarity 0.84)
- VISTA3D: A Unified Segmentation Foundation Model For 3D Medical Imaging (CVPR 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.