Research map / Detection and segmentation / Weakly supervised detection and segmentation
Weakly Supervised Gland Segmentation with Class Semantic Consistency and Purified Labels Filtration
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 CCA, MPFP and PLF modules for image-level weakly supervised gland segmentation from CAMs, reporting new SOTA mIoU on ProG, GlaS and EBHI.
Problem
CAM-based image-level weakly supervised segmentation performs poorly on histopathological gland images due to low tissue contrast, morphological homogeneity and class/semantic confusion.
Approach
Combines a class-consistency attention (CCA, criss-cross plus position attention) inside the classifier, a multi-scale pyramid fusion pooling (MPFP) with sign-preserving convolutions to replace GAP for fine-grained CAMs, and a purified labels filtration (PLF) loss that down-weights noisy pixels during segmentation training.
Key results
- Achieves 87.69% mIoU on ProG, 80.44% on GlaS and 77.36% on EBHI, the best among compared weakly supervised methods.
- Ablation on ProG: baseline 80.01% mIoU rises to 87.69% with CCA, MPFP and PLF combined.
- Domain adaptation: fine-tuning with only 1% of target data keeps stable performance where OEEM drops substantially.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the full method with CCA, MPFP, and PLF | ResNet38; PSPNet with ResNeSt200 · ProG | mIoU: 84.56±0.13 -> 87.69±0.12 | Helped |
| Use the full method with CCA, MPFP, and PLF | ResNet38; PSPNet with ResNeSt200 · ProG | F1-score: 91.63±0.07 -> 93.44±0.09 | Helped |
| Use the full method with CCA, MPFP, and PLF | ResNet38; PSPNet with ResNeSt200 · GlaS | mIoU: 77.37±0.73 -> 80.44±0.05 | Helped |
| Use the full method with CCA, MPFP, and PLF | ResNet38; PSPNet with ResNeSt200 · GlaS | F1-score: 87.24±0.46 -> 89.16±0.03 | Helped |
| Use the full method with CCA, MPFP, and PLF | ResNet38; PSPNet with ResNeSt200 · EBHI | mIoU: 75.79±0.31 -> 77.36±0.23 | Helped |
| Use the full method with CCA, MPFP, and PLF | ResNet38; PSPNet with ResNeSt200 · EBHI | F1-score: 86.26±0.17 -> 87.24±0.11 | Helped |
| Add CCA (ablation) | Baseline gland segmentation method · ProG | mIoU: 80.01±0.46 -> 82.90±0.10 | Helped |
| Add MPFP (ablation) | Baseline gland segmentation method · ProG | mIoU: 80.01±0.46 -> 86.11±0.15 | Helped |
| Add PLF to CCA + MPFP (ablation) | CCA + MPFP gland segmentation method · ProG | mIoU: 86.82±0.08 -> 87.69±0.12 | Helped |
| Apply CCA at Stages 4, 5, and 6 rather than at one stage only (ablation) | ResNet38 classification network · ProG | mIoU: 85.25±0.14; 86.43±0.14; 86.91±0.15 -> 87.69±0.12 | Helped |
| Use CCA instead of one or two criss-cross attention blocks (ablation) | ResNet38 classification network · ProG | mIoU: 84.88±0.21; 86.34±0.23 -> 87.69±0.12 | Helped |
| Fuse multiple pooling scales with sign-preserving convolution rather than using a single scale or average fusion (ablation) | MPFP · ProG | mIoU: 83.52±0.17; 84.79±0.15; 85.86±0.15; 86.77±0.18; 87.10±0.10 -> 87.69±0.12 | Helped |
Surprising findings
- On GlaS and EBHI the weakly supervised method's mIoU is reported as close to or comparable with the fully supervised baseline.
Limitations
- PLF is applied only to multi-class images, not single-class ones.
- Evaluated on three gland datasets; generalization beyond gland segmentation is not shown.
Experimental setup
Datasets: ProG GlaS EBHI
Benchmarks: ProG GlaS EBHI
Metrics: mIoU F1-score p-value (independent samples t-test)
Baselines: SEAM ReCAM AMR MLPS OEEM AME-CAM HAMIL CBFNet
Builds on: CAM criss-cross attention PSPNet SEAM
Closest papers by meaning
- Prototype-Based Image Prompting for Weakly Supervised Histopathological Image Segmentation (CVPR 2025 · similarity 0.88)
- W2P: Switching from Weak Supervision to Partial Supervision for Semantic Segmentation (AAAI 2024 · similarity 0.87)
- MATCH: Multi-faceted Adaptive Topo-Consistency for Semi-Supervised Histopathology Segmentation (NeurIPS 2025 · similarity 0.86)
- Class Tokens Infusion for Weakly Supervised Semantic Segmentation (CVPR 2024 · similarity 0.86)
- DuPL: Dual Student with Trustworthy Progressive Learning for Robust Weakly Supervised Semantic Segmentation (CVPR 2024 · similarity 0.86)
- Progressive Feature Self-Reinforcement for Weakly Supervised Semantic Segmentation (AAAI 2024 · similarity 0.86)
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.