Research map / Continual learning / Class-incremental
Incremental Nuclei Segmentation from Histopathological Images via Future-class Awareness and Compatibility-inspired Distillation
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 future-class awareness and compatibility-inspired distillation for class-incremental nuclei segmentation from histopathology images without old-class data, evaluated on MoNuSAC and CoNSeP.
Problem
Class-incremental semantic segmentation of cell nuclei in histopathology images where old-class training data is unavailable, so models catastrophically forget old classes while needing to learn new ones.
Approach
Unsupervisedly detects potential future-class pixels in background by affinity to old/new class centroids in the old model's representation space and labels them as unknown; then expands the old model's output space with zero-padded channels and transfers background/future predictions to new-class ground truth (compatibility-inspired distillation), building on EWF weight fusion and Local POD.
Key results
- On MoNuSAC 1-1 (4 tasks), the method reaches 69.44 mean mDice vs. 67.06 for EWF and 67.28 for CoNuSeg.
- On MoNuSAC 3-1, mean mDice is 72.36 vs. 71.20 for EWF.
- On CoNSeP 2-1, mean mDice is 71.08 vs. 70.44 for EWF.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add future-class awareness to PLOP + EWF (ablation) | Incremental nuclei segmentation model · MoNuSAC 1-1 (4 tasks) | mDice (Mean): 67.06 -> 68.60 | Helped |
| Add future-class awareness to PLOP + EWF (ablation) | Incremental nuclei segmentation model · MoNuSAC 3-1 (2 tasks) | mDice (Mean): 71.20 -> 72.07 | Helped |
| Add compatibility-inspired distillation to PLOP + EWF + FCA (ablation) | Incremental nuclei segmentation model · MoNuSAC 1-1 (4 tasks) | mDice (Mean): 68.60 -> 69.42 | Helped |
| Add compatibility-inspired distillation to PLOP + EWF + FCA (ablation) | Incremental nuclei segmentation model · MoNuSAC 3-1 (2 tasks) | mDice (Mean): 72.07 -> 72.36 | Helped |
| Set αn=1 and αb=1 in CID instead of αn=0 and αb=0 (ablation) | Compatibility-inspired distillation · MoNuSAC 1-1 (4 tasks) | All: 73.80 -> 74.03 | Helped |
| Set αn=3 and αb=1 in CID instead of αn=1 and αb=1 (ablation) | Compatibility-inspired distillation · MoNuSAC 1-1 (4 tasks) | All: 74.03 -> 74.42 | Helped |
| Set αn=5 and αb=1 in CID instead of αn=3 and αb=1 (ablation) | Compatibility-inspired distillation · MoNuSAC 1-1 (4 tasks) | Old: 68.11 -> 65.91 | Hurt |
| Set αn=3 and αb=3 in CID instead of αn=3 and αb=1 (ablation) | Compatibility-inspired distillation · MoNuSAC 1-1 (4 tasks) | New: 69.86 -> 68.34 | Hurt |
| Use the proposed method instead of EWF | Incremental nuclei segmentation model · MoNuSAC 1-1 (4 tasks) | mDice (Mean): 67.06 -> 69.44 | Helped |
| Use the proposed method instead of EWF | Incremental nuclei segmentation model · MoNuSAC 2-1 (3 tasks) | mDice (Mean): 68.26 -> 70.16 | Helped |
| Use the proposed method instead of EWF | Incremental nuclei segmentation model · MoNuSAC 2-2 (2 tasks) | mDice (Mean): 70.92 -> 71.32 | Helped |
| Use the proposed method instead of EWF | Incremental nuclei segmentation model · MoNuSAC 3-1 (2 tasks) | mDice (Mean): 71.20 -> 72.36 | Helped |
Limitations
- May not perform well when future classes differ visually from learned classes.
- Gains on new classes are modest in the most challenging 1-1 setting (0.45% over IDEC).
Experimental setup
Datasets: MoNuSAC CoNSeP
Benchmarks: MoNuSAC CoNSeP
Metrics: mDice
Baselines: ILT MiB SDR PLOP REMINDER CoNuSeg IDEC EWF
Builds on: PLOP EWF Local POD MiB
Closest papers by meaning
- Prototype-Based Image Prompting for Weakly Supervised Histopathological Image Segmentation (CVPR 2025 · similarity 0.87)
- MATCH: Multi-faceted Adaptive Topo-Consistency for Semi-Supervised Histopathology Segmentation (NeurIPS 2025 · similarity 0.86)
- UniCell: Universal Cell Nucleus Classification via Prompt Learning (AAAI 2024 · similarity 0.85)
- Weakly Supervised Gland Segmentation with Class Semantic Consistency and Purified Labels Filtration (AAAI 2025 · 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 0.85)
- FOCUS: Knowledge-enhanced Adaptive Visual Compression for Few-shot Whole Slide Image Classification (CVPR 2025 · similarity 0.84)
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.