Research map / Detection and segmentation / CNN features
Category Prompt Mamba Network for Nuclei Segmentation and Classification
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 CP-Mamba, a Mamba-based encoder with category-prompt supervision and probability-guided sequence sorting that trains nuclei segmentation and classification directly on full-sized images without patch splitting.
Problem
Nuclei segmentation and classification models split large pathology images into patches, causing border misalignment and slow training/inference, and class imbalance degrades rare-category performance.
Approach
Category prompt supervision via patch-level multi-label targets produces per-class probabilities; feature sequences are sorted per category by descending confidence before Mamba scanning, so low-confidence features learn from high-confidence ones, enabling direct full-image training.
Key results
- On CoNSeP, the method reaches DICE 82.23, AJI 53.90, PQ 51.46 for segmentation, outperforming HoverNet and Vim.
- Ablation on GlySAC: probability sorting plus phenotype learning raises F_d from 79.42 to 86.94 and Dice from 79.27 to 81.43.
- Classification F1 on ConSeP miscellaneous nuclei improves by more than 20 points over DIST and DSF-CNN (e.g., 40.96 vs 17.18 for DIST).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use probability sorting instead of bidirectional scanning (ablation) | CP-Mamba · GlySAC | miscellaneous nuclei F1: 29.27 -> 35.92 | Helped |
| use probability sorting instead of cross-scanning (ablation) | CP-Mamba · GlySAC | miscellaneous nuclei F1: 31.08 -> 35.92 | Helped |
| use probability sorting instead of bidirectional scanning (ablation) | CP-Mamba · ConSeP | miscellaneous nuclei F1: 25.33 -> 40.96 | Helped |
| use probability sorting instead of cross-scanning (ablation) | CP-Mamba · ConSeP | miscellaneous nuclei F1: 34.13 -> 40.96 | Helped |
| add phenotype learning (ablation) | CP-Mamba · GlySAC | detection F1 (Fd): 80.20 -> 86.94 | Helped |
| add phenotype learning (ablation) | CP-Mamba · ConSeP | detection F1 (Fd): 72.22 -> 75.33 | Helped |
| add probability sorting (ablation) | CP-Mamba · GlySAC | detection F1 (Fd): 84.75 -> 86.94 | Helped |
| add probability sorting (ablation) | CP-Mamba · ConSeP | detection F1 (Fd): 73.83 -> 75.33 | Helped |
| add phenotype learning and probability sorting (ablation) | CP-Mamba · GlySAC | detection F1 (Fd): 79.42 -> 86.94 | Helped |
| add phenotype learning and probability sorting (ablation) | CP-Mamba · ConSeP | detection F1 (Fd): 71.12 -> 75.33 | Helped |
Limitations
- Theoretical guarantee for sorting is stated as a theorem with proof deferred to supplementary material, which is not in the provided text.
- Relies on hand-set hyperparameters such as lambda=0.2 and alpha=beta=1 without broad tuning in the main text.
- Improvements on some categories (e.g., PanNuke neoplastic-related and CoNSeP epithelial) are marginal or absent.
Experimental setup
Datasets: GLySAC CoNSeP MoNuSAC PanNuke
Benchmarks: GLySAC CoNSeP MoNuSAC PanNuke
Metrics: DICE AJI DQ SQ PQ F1
Baselines: Mask-RCNN HoverNet Triple-UNet DoNet Vim DIST DSF-CNN SONNET CGT SENC
Builds on: Mamba VMamba Vim HoverNet
Closest papers by meaning
- S³-Mamba: Small-Size-Sensitive Mamba for Lesion Segmentation (AAAI 2025 · similarity 0.85)
- Incremental Nuclei Segmentation from Histopathological Images via Future-class Awareness and Compatibility-inspired Distillation (CVPR 2024 · similarity 0.82)
- ZePT: Zero-Shot Pan-Tumor Segmentation via Query-Disentangling and Self-Prompting (CVPR 2024 · similarity 0.82)
- Cross-Modal Interactive Perception Network with Mamba for Lung Tumor Segmentation in PET-CT Images (CVPR 2025 · similarity 0.81)
- MambaOut: Do We Really Need Mamba for Vision? (CVPR 2025 · similarity 0.81)
- Hybrid Mamba for Few-Shot Segmentation (NeurIPS 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.