Research map / Detection and segmentation / Open-set and few-shot detection
Diversified and Personalized Multi-rater Medical Image Segmentation
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 D-Persona, a two-stage multi-rater medical segmentation framework that first learns a diversified Probabilistic U-Net latent space with a bound-constrained loss, then uses attention-based projection heads to query expert-specific prompts for personalized segmentation.
Problem
Annotation ambiguity from data uncertainty and observer preferences in medical image segmentation; prior methods either produce a consensus, diverse implicit outputs, or personalized per-expert outputs, but not both.
Approach
Two-stage design: Stage I trains a Probabilistic U-Net with an added intersection/union bound loss on prior samples to increase diversity and build a shared latent space; Stage II trains per-expert projection heads whose pooled codes are mapped by cross-attention onto a frozen prior bank, yielding personalized segmentations from the same latent space.
Key results
- On NPC-170, D-Persona Stage II reaches Dice_mean 76.19% versus 75.09% for Pionono and 72.99% for majority voting.
- Stage I lowers GED to 0.2359 on NPC-170 (vs 0.4465 for Prob. U-Net with 50 samples) and raises soft Dice to 80.43%.
- On LIDC-IDRI, Stage II achieves Dice_mean 89.17% vs 88.84% for Pionono, and Stage I reaches GED 0.1358 vs 0.2168 for Prob. U-Net.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the bound-constrained loss to Probabilistic U-Net | D-Persona Stage I · NPC-170 | GED: 0.4465 -> 0.2359 | Helped |
| Add the bound-constrained loss to Probabilistic U-Net | D-Persona Stage I · LIDC-IDRI | Dicesoft: 88.80 -> 90.45 | Helped |
| Use two-stage personalization with attention-based projection heads | D-Persona · NPC-170 | Dicemean: 75.09 -> 76.19; a 1.10% gain | Helped |
| Use two-stage personalization with attention-based projection heads | D-Persona · LIDC-IDRI | Dicemean: 88.84 -> 89.17 | Helped |
| Increase the number of sampled latent codes from #10 to #50 | D-Persona Stage I · NPC-170 | Dicemax: a 1.16% gain | Helped |
| Set K to 10 rather than the other tested values (ablation) | D-Persona Stage I · NPC-170 | Dicesoft: 78.81, 79.48, 80.43, 80.17, 80.08 | Helped |
| Set β to 0.5 rather than the other tested values (ablation) | D-Persona Stage I · NPC-170 | Dicesoft: 75.80, 77.95, 80.43, 80.20, 77.73 | Helped |
| Set β to 1 instead of β=0.5 (ablation) | D-Persona Stage I · NPC-170 | GED: 0.2359 -> 0.2170 | Helped |
Surprising findings
- Stage II does not beat individually trained single-expert U-Nets on personalized Dice for most experts on NPC-170, with only expert A4 improved.
Limitations
- Stage II personalized results on NPC-170 do not surpass individually trained U-Nets for most experts.
- Evaluated only on two limited datasets, with possible dataset bias.
- LIDC-IDRI expert preferences are simulated by ranking annotation areas rather than using real expert identities.
Experimental setup
Datasets: NPC-170 LIDC-IDRI
Benchmarks: NPC-170 LIDC-IDRI
Metrics: GED soft Dice Dice max Dice match per-expert Dice mean Dice
Baselines: U-Net (single expert) Prob. U-Net CM-Global CM-Pixel TAB Pionono majority voting random selection STAPLE
Builds on: Probabilistic U-Net Pionono STAPLE random selection for multi-rater labels
Closest papers by meaning
- Versatile Medical Image Segmentation Learned from Multi-Source Datasets via Model Self-Disambiguation (CVPR 2024 · similarity 0.87)
- Annotation Ambiguity Aware Semi-Supervised Medical Image Segmentation (CVPR 2025 · similarity 0.86)
- Enhancing SAM with Efficient Prompting and Preference Optimization for Semi-supervised Medical Image Segmentation (CVPR 2025 · similarity 0.86)
- Mamba Goes HoME: Hierarchical Soft Mixture-of-Experts for 3D Medical Image Segmentation (NeurIPS 2025 · similarity 0.86)
- Distribution-aware Fairness Learning in Medical Image Segmentation From A Control-Theoretic Perspective (ICML 2025 · similarity 0.85)
- SGTC: Semantic-Guided Triplet Co-training for Sparsely Annotated Semi-Supervised Medical Image Segmentation (AAAI 2025 · similarity 0.85)
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.