Research map / Detection and segmentation / CNN features
Every Component Counts: Rethinking the Measure of Success for Medical Semantic Segmentation in Multi-Instance Segmentation Tasks
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 CC-Metrics, which partitions a volume into generalized Voronoi regions around ground-truth connected components and computes standard segmentation metrics per component, to remove size bias in multi-instance medical segmentation evaluation on PET/CT.
Problem
Global semantic segmentation metrics (Dice, Surface Dice, Hausdorff distances, Panoptic Quality, Lesion Dice) are biased towards large connected components, so they misrepresent performance on multi-instance tumor/metastasis segmentation where each lesion should count equally.
Approach
Evaluate standard overlap, boundary, and distance metrics locally per connected component within generalized Voronoi regions (proximity-based matching, no IoU threshold), then average across components, giving every metastasis equal weight regardless of size.
Key results
- On AutoPET, nnUNet achieves Dice 67.5 but CC-Dice 47.4, showing large gaps for small metastases.
- CC-Metrics rank DynUNet above UNETR on CC-Dice (30.5 vs 28.0), reversing the Dice ranking (33.5 vs 41.6).
- On HECKTOR (about 2 metastases per patient), CC-Dice and Dice are nearly identical (e.g., DynUnet 78.7 vs 78.7), indicating alignment when component counts are low.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Evaluate Dice per connected component and average across components instead of scoring globally | Segmentation models on AutoPET · AutoPET | Dice: 67.5 -> 47.4; 33.5 -> 30.5; 41.6 -> 28.0; 54.8 -> 38.4 | Mixed |
| Evaluate Surface Dice per connected component and average across components instead of scoring globally | Segmentation models on AutoPET · AutoPET | Surface Dice: 66.3 -> 52.5; 28.5 -> 32.8; 33.9 -> 29.3; 49.0 -> 40.2 | Mixed |
| Evaluate Hausdorff distance per connected component and average across components instead of scoring globally | Segmentation models on AutoPET · AutoPET | HD: 71.1 -> 46.4; 190 -> 114; 211 -> 157; 174 -> 130 | Helped |
| Evaluate Dice per connected component and average across components instead of scoring globally | Segmentation models on HECKTOR · HECKTOR | Dice: 78.7 -> 78.7; 60.8 -> 60.7; 72.3 -> 72.3 | No clear effect |
| Evaluate Surface Dice per connected component and average across components instead of scoring globally | Segmentation models on HECKTOR · HECKTOR | Surface Dice: 80.3 -> 80.3; 56.8 -> 56.8; 71.2 -> 71.3 | No clear effect |
| Evaluate Hausdorff distance per connected component and average across components instead of scoring globally | Segmentation models on HECKTOR · HECKTOR | HD: 37.7 -> 37.6; 151 -> 151; 113 -> 113 | No clear effect |
| Drop the 10 smallest metastases in synthetic predictions | Standard Dice and CC-Dice evaluation · AutoPET | Dice | Mixed |
| Drop small or large components and score each component separately | CC-Dice and CC-Surface Dice evaluation · AutoPET | Dice | Helped |
| Add false-positive components to selected regions | CC-Dice and CC-Surface Dice evaluation · AutoPET | Dice | Helped |
| Use CC-HD95 to average per-component worst-case distances rather than a global worst case | Hausdorff distance evaluation · AutoPET | HD95 | Helped |
| Use per-component distance evaluation when eroding the smallest rather than largest component | Hausdorff95 evaluation · Synthetic component erosion | HD95 | Helped |
| Use per-component Surface Dice instead of global Surface Dice when shifting a small component | Surface Dice evaluation · Synthetic component shift | Surface Dice | Helped |
Surprising findings
- Standard Dice barely changes when the 10 smallest metastases are dropped, while CC-Dice reflects the degradation.
- Panoptic Quality and Lesion Dice can increase when a false-positive component is fully eroded, rewarding absence of predictions.
Limitations
- Proximity-based matching assumes constant pixel spacing across the domain (e.g., MRI or CT).
- Evaluated on only two PET/CT datasets and four segmentation architectures, with models not tuned for performance.
- Per-component averaging gives equal weight to all components, which may not match every clinical priority.
Experimental setup
Datasets: AutoPET-II HECKTOR
Benchmarks: AutoPET-II HECKTOR
Metrics: Dice CC-Dice Surface Dice CC-Surface Dice Hausdorff95 CC-HD95 Panoptic Quality Lesion Dice
Baselines: Dice Surface Dice Hausdorff95 Panoptic Quality Lesion Dice
Builds on: Panoptic Quality Lesion Dice (BraTS 2023) nnU-Net MONAI
Closest papers by meaning
- S2S2: Semantic Stacking for Robust Semantic Segmentation in Medical Imaging (AAAI 2025 · similarity 0.84)
- Weakly Supervised Gland Segmentation with Class Semantic Consistency and Purified Labels Filtration (AAAI 2025 · similarity 0.84)
- UniPro: Unified Multi-Mode Medical Image Segmentation from 2D Images to 3D Volumes via Propagation (NeurIPS 2026 · similarity 0.84)
- Rethinking U-Net: Task-Adaptive Mixture of Skip Connections for Enhanced Medical Image Segmentation (AAAI 2025 · similarity 0.84)
- Pancakes: Consistent Multi-Protocol Image Segmentation Across Biomedical Domains (NeurIPS 2025 · similarity 0.84)
- CAT: Coordinating Anatomical-Textual Prompts for Multi-Organ and Tumor Segmentation (NeurIPS 2024 · similarity 0.83)
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.