Research map / Detection and segmentation / CNN features
The Four Color Theorem for Cell Instance 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.
Reformulates cell instance segmentation as a four-class semantic segmentation using a greedy four-color encoding of touching cells, with an asymptotic training strategy, sampling-based orthogonality loss and an encoding transformation module.
Problem
Accurately separating tightly touching cells in biomedical images while keeping model complexity and post-processing low; existing detection-, contour- and distance-based instance segmentation methods are either complex or sensitive to thresholds.
Approach
Encode each cell's instance label with a four-color greedy coloring so adjacent cells get distinct colors, turning instance segmentation into four-class semantic segmentation; stabilize training via a foreground/background-then-color asymptotic strategy, a negative-sampling cosine orthogonality loss between adjacent cells, and a learned encoding transformation that maps predictions to a canonical
Key results
- On DSB2018, FCIS reaches DICE 0.939, AJI 0.828, PQ 0.770, the best among compared methods.
- On PanNuke, FCIS achieves PQ 0.610 and AJI 0.653, outperforming all compared methods.
- FCIS uses 39.75M parameters and 58.03G FLOPs at 256x256 input, fewer than HoverNet (49.70M, 192.70G) and CellViT, though CellPose and others differ in cost.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add asymptotic training and encoding transformation (ablation) | FCIS · DSB2018 | DICE: 0.862 -> 0.883 | Helped |
| add asymptotic training and encoding transformation (ablation) | FCIS · PanNuke | DICE: 0.773 -> 0.787 | Helped |
| add negative sampling constraint for adjacent cells (ablation) | FCIS · DSB2018 | DICE: 0.883 -> 0.939 | Helped |
| add negative sampling constraint for adjacent cells (ablation) | FCIS · PanNuke | DICE: 0.787 -> 0.816 | Helped |
| directly use four-color encoding as supervision (ablation) | FCIS · DSB2018 | DICE: 0.876 -> 0.843(-3.3) | Hurt |
| directly use four-color encoding as supervision (ablation) | FCIS · PanNuke | DICE: 0.786 -> 0.766(-2.0) | Hurt |
| increase sampling rate from 0.3 to 0.5 (ablation) | FCIS · DSB2018 | DICE: 0.913 -> 0.939 | Helped |
| increase sampling rate from 0.5 to 0.7 (ablation) | FCIS · DSB2018 | DICE: 0.939 -> 0.941 | Helped |
| increase sampling rate from 0.3 to 0.5 (ablation) | FCIS · BBBC006v1 | DICE: 0.947 -> 0.954 | Helped |
| increase sampling rate from 0.5 to 0.7 (ablation) | FCIS · BBBC006v1 | DICE: 0.954 -> 0.946 | Hurt |
| set orthogonal constraint loss weight from 2 to 1 (ablation) | FCIS · DSB2018 | DICE: 0.939 -> 0.908 | Hurt |
| set orthogonal constraint loss weight from 2 to 1 (ablation) | FCIS · BBBC006v1 | DICE: 0.954 -> 0.922 | Hurt |
Surprising findings
- Directly supervising with the four-color encoding drops DSB2018 PQ from 0.746 to 0.674, showing naive encoding supervision hurts training.
- The supplementary analysis reports that in the datasets the fourth color is never used, suggesting cell colorings are simpler than map colorings.
Limitations
- The theoretical guarantees (Theorem 1 and 2) rely on informal arguments, and the greedy optimality claim for planar cell graphs is not rigorously proven.
- Reported gains are on a single split per dataset with no variance estimates.
- Adaptive encoding for varying tissue architectures and cell densities is left to future work.
Experimental setup
Datasets: BBBC006v1 DSB2018 PanNuke YeaZ
Benchmarks: DSB2018 PanNuke BBBC006v1 YeaZ
Metrics: DICE AJI DQ SQ PQ
Baselines: DoNet DCAN NucleiSegNet GeNSeg-Net HoverNet CPP-Net CellPose Un-SAM
Builds on: U-Net four-color theorem Dice loss cross-entropy loss
Closest papers by meaning
- Disco: Densely-overlapping Cell Instance Segmentation via Adjacency-aware Collaborative Coloring (ICLR 2026 · similarity 0.87)
- Object-Guided Instance Segmentation for Biological Images (AAAI 2020 · similarity 0.86)
- Weakly Supervised Gland Segmentation with Class Semantic Consistency and Purified Labels Filtration (AAAI 2025 · similarity 0.82)
- P2Seg: Pointly-supervised Segmentation via Mutual Distillation (ICLR 2024 · similarity 0.82)
- FISBe: A Real-World Benchmark Dataset for Instance Segmentation of Long-Range Thin Filamentous Structures (CVPR 2024 · similarity 0.81)
- PolarNeXt: Rethink Instance Segmentation with Polar Representation (CVPR 2025 · similarity 0.81)
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.