Research map / Detection and segmentation / Weakly supervised detection and segmentation
MonoBox: Tightness-Free Box-Supervised Polyp Segmentation Using Monotonicity Constraint
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 MonoBox, a tightness-free box-supervised polyp segmentation method that replaces MIL loss in noisy box-edge zones with a monotonicity constraint and shrinks those zones via label correction.
Problem
Box-supervised polyp segmentation relies on the box-tightness assumption (box edges touch target boundaries), which fails for polyps with vague boundaries, producing noisy MIL supervision.
Approach
Within unconfident zones near box edges, discard MIL loss and enforce a monotonicity constraint (inner response higher than outer response along directions toward the box center), plus adaptive label correction using predicted-mask boxes that shrinks the unconfident zones during training.
Key results
- On the synthetic noisy dataset, IBoxCLA with MonoBox reaches Dice 0.803 vs 0.735 for the noisy-label baseline (LB) and 0.827 for the clean upper bound (UB), beating FSRM (0.748) by 5.5%.
- On the in-house real noisy dataset, IBoxCLA with MonoBox reaches Dice 0.849 vs 0.816 for FSRM, a 3.3% gain.
- On COCO with BoxInst, MonoBox raises AP from 0.248 (LB) to 0.308, close to UB 0.321, and beats FSRM (0.259).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply MonoBox instead of FSRM | WeakPolyp · Real noisy | Dice: 0.785 -> 0.804 | Helped |
| apply MonoBox instead of FSRM | WeakPolyp · Synthetic noisy | Dice: 0.722 -> 0.763 | Helped |
| apply MonoBox instead of FSRM | IBoxCLA · Real noisy | Dice: 0.816 -> 0.849 | Helped |
| apply MonoBox instead of FSRM | IBoxCLA · Synthetic noisy | Dice: 0.748 -> 0.803 | Helped |
| add monotonicity constraint (ablation) | IBoxCLA box-supervised segmentation · Real noisy | Dice: 0.803 -> 0.838 | Helped |
| add monotonicity constraint (ablation) | IBoxCLA box-supervised segmentation · Synthetic noisy | Dice: 0.735 -> 0.787 | Helped |
| add label correction without MC (ablation) | IBoxCLA box-supervised segmentation · Real noisy | Dice: 0.803 -> 0.809 | Helped |
| add label correction without MC (ablation) | IBoxCLA box-supervised segmentation · Synthetic noisy | Dice: 0.735 -> 0.744 | Helped |
| combine monotonicity constraint and label correction (ablation) | IBoxCLA box-supervised segmentation · Real noisy | Dice: 0.838 -> 0.849 | Helped |
| combine monotonicity constraint and label correction (ablation) | IBoxCLA box-supervised segmentation · Synthetic noisy | Dice: 0.787 -> 0.803 | Helped |
| exclude the unconfident region from loss calculation instead of using MC (ablation) | IBoxCLA box-supervised segmentation · Real noisy | Dice: 0.803 -> 0.809 | No clear effect |
| use 2D-Gaussian soft labels instead of MC (ablation) | IBoxCLA box-supervised segmentation · Real noisy | Dice: 0.803 -> 0.816 | Helped |
Surprising findings
- Using label correction alone yields only limited improvement; its benefit appears mainly when combined with the monotonicity constraint.
- Simply excluding unconfident regions from the loss hardly improves over the noisy-label baseline.
Limitations
- Uses a uniform unconfident scale for all training samples rather than per-sample adaptation.
- Synthetic noise is zero-mean Gaussian and may not reflect real annotator bias (e.g., over-sized boxes).
Experimental setup
Datasets: ClinicDB Kvasir-SEG ColonDB EndoScene ETIS in-house polyp dataset COCO
Benchmarks: ClinicDB Kvasir-SEG ColonDB EndoScene ETIS in-house real noisy polyp test set COCO test-dev
Metrics: Dice IoU Hausdorff distance AP
Baselines: SSD-Det NCE+RCE PolarT FSRM Soft Label (2D-Gaussian) Exclusion BoxInst WeakPolyp IBoxCLA
Builds on: multiple-instance learning BoxInst IBoxCLA FSRM
Closest papers by meaning
- Polyper: Boundary Sensitive Polyp Segmentation (AAAI 2024 · similarity 0.84)
- Boost the Inference with Co-training: A Depth-guided Mutual Learning Framework for Semi-supervised Medical Polyp Segmentation (CVPR 2025 · similarity 0.83)
- Why Invariance is Not Enough for Biomedical Domain Generalization and How to Fix It (NeurIPS 2026 · similarity 0.83)
- Sketchy Bounding-box Supervision for 3D Instance Segmentation (CVPR 2025 · similarity 0.82)
- WISH: Weakly Supervised Instance Segmentation using Heterogeneous Labels (CVPR 2025 · similarity 0.82)
- Weakly Supervised Gland Segmentation with Class Semantic Consistency and Purified Labels Filtration (AAAI 2025 · similarity 0.82)
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.