Research map / Detection and segmentation / Weakly supervised detection and segmentation
Mudslide: A Universal Nuclear Instance Segmentation Method
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 Mudslide, a collapse-field-based nuclear instance segmentation method inspired by mudslides that separates overlapping nuclei by pixel-wise collapse along force directions, and extends to point supervision.
Problem
Nuclear instance segmentation in pathology struggles with densely overlapping nuclei and blurry boundaries, and high annotation cost limits mask-level supervision; existing boundary-based methods fail to differentiate overlapping instances and do not transfer well to point supervision.
Approach
Builds a four-channel collapse field (distance-to-center maps in horizontal, vertical, and diagonal directions), derives an eight-direction force map and initial boundary (Sobel edges) from it, and assigns instances by pixel-by-pixel collapse along the force directions, so pixels collapsing into the same region form one instance; the same pipeline is used for full and point supervision.
Key results
- Fully supervised: 0.714 PQ and 0.740 AJI on CPM17, and 0.599 PQ and 0.647 AJI on NuInsSeg, slightly above CellViT (0.713 PQ on CPM17, 0.596 PQ on NuInsSeg).
- Point supervised: 0.529 PQ and 0.581 AJI on CPM17 and 0.456 PQ and 0.558 AJI on NuInsSeg, outperforming SCNet (0.496 PQ on CPM17, 0.456 PQ on NuInsSeg).
- Ablation: initial boundary from collapse field plus mudslide collapse improves PQ from 0.548 to 0.714 on CPM17 over a connected-components baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Mudslide instead of the strongest listed fully supervised baseline | Trans-UNet nuclear instance segmentation · CPM17 | PQ: 0.713 -> 0.714 | Helped |
| Use Mudslide instead of the strongest listed fully supervised baseline | Trans-UNet nuclear instance segmentation · NuInsSeg | AJI: 0.640 -> 0.647 | Helped |
| Use Mudslide instead of the strongest listed point-supervised baseline | Trans-UNet nuclear instance segmentation · CPM17 | PQ: 0.496 -> 0.529 | Helped |
| Use Mudslide instead of the strongest listed point-supervised baseline | Trans-UNet nuclear instance segmentation · NuInsSeg | AJI: 0.526 -> 0.558 | Helped |
| Add an initial boundary to binary-map connected-region partitioning (ablation) | Mudslide algorithm · CPM17 | PQ: 0.548 -> 0.707 | Helped |
| Add an initial boundary to binary-map connected-region partitioning (ablation) | Mudslide algorithm · NuInsSeg | PQ: 0.414 -> 0.582 | Helped |
| Add mudslide collapse after using the initial boundary (ablation) | Mudslide algorithm · CPM17 | PQ: 0.707 -> 0.714 | Helped |
| Add mudslide collapse after using the initial boundary (ablation) | Mudslide algorithm · NuInsSeg | PQ: 0.582 -> 0.599 | Helped |
| Extract force directions from a four-channel collapse field instead of direct force-map prediction (ablation) | Mudslide algorithm · CPM17 | PQ: 0.707 -> 0.711 | Helped |
| Extract force directions from a four-channel collapse field instead of two-channel collapse-field segmentation (ablation) | Mudslide algorithm · NuInsSeg | AJI: 0.629 -> 0.643 | Helped |
| Generate the initial boundary from the collapse field instead of the binary prediction map (ablation) | Mudslide algorithm · CPM17 | PQ: 0.711 -> 0.714 | Helped |
| Generate the initial boundary from the collapse field instead of the binary prediction map (ablation) | Mudslide algorithm · NuInsSeg | PQ: 0.592 -> 0.599 | Helped |
Limitations
- Fully supervised gains over CellViT are marginal (about 0.001 to 0.003 PQ), so the claimed superiority is small.
- Point-supervised performance still lags far behind fully supervised results.
- Evaluated only on two H&E nuclear datasets with single-class segmentation.
Experimental setup
Datasets: CPM17 NuInsSeg
Benchmarks: CPM17 NuInsSeg
Metrics: DQ SQ PQ AJI
Baselines: Watershed Boundary partition HoverNet CellViT UNet DCAN DIST FullNet CDNet PENet C2FNet S2Label
Builds on: Hover-Net CellViT U-Net
Closest papers by meaning
- Supervise Less, See More: Training-free Nuclear Instance Segmentation with Prototype-Guided Prompting (ICML 2026 · similarity 0.87)
- Minimizing Labeling Cost for Nuclei Instance Segmentation and Classification with Cross-domain Images and Weak Labels (AAAI 2021 · similarity 0.85)
- Are nuclear masks all you need for improved out-of-domain generalisation? A closer look at cancer classification in histopathology (NeurIPS 2024 · similarity 0.83)
- PolarNeXt: Rethink Instance Segmentation with Polar Representation (CVPR 2025 · similarity 0.82)
- PathVQ: Reforming Computational Pathology Foundation Model for Whole Slide Image Analysis via Vector Quantization (NeurIPS 2025 · similarity 0.82)
- Incremental Nuclei Segmentation from Histopathological Images via Future-class Awareness and Compatibility-inspired Distillation (CVPR 2024 · 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.