Research map / Label-efficient and robust learning / Noisy labels
Minding Fuzzy Regions: A Data-driven Alternating Learning Paradigm for Stable Lesion Segmentation
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 DALE, a model-agnostic alternating training paradigm that splits medical images into fuzzy and non-fuzzy regions, learns per-label confidence via loss-consistency optimization, and aligns fuzzy-region feature distributions to non-fuzzy ones with a Wasserstein loss, improving lesion segmentation across several backbones.
Problem
Lesion segmentation in medical images suffers from unclear boundaries, irregular shapes and low tissue contrast, which produce noisy labels and unstable feature representations in fuzzy regions that standard training treats equally with clean data.
Approach
Alternate training between reliable non-fuzzy regions and fuzzy regions; in fuzzy regions, learn label confidence weights by gradient/loss consistency with a shadow model, and calibrate fuzzy-region features by aligning their per-class Gaussian distributions to those of non-fuzzy regions via a Wasserstein distance loss.
Key results
- DALE improves Dice across 9 backbones on MosMedData+ by an average of 5.33%, reaching 0.780 Dice for LanGuide (from 0.735).
- On the Polyp datasets DALE gives an average Dice gain of 7.16%; on ISIC2016&ph2 the average gain is 2.37%.
- On ISIC2016&ph2 with Xboundformer, DALE reaches Dice 0.936 vs 0.917 for the base model and 0.928 for CIRL.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply DALE instead of CIRL | Xboundformer · ISIC2016&ph2 | Dice: 0.928 -> 0.936 | Helped |
| apply DALE instead of CIRL | Xboundformer · ISIC2016&ph2 | 95HD: 4.097 -> 3.224 | Helped |
| apply DALE instead of CIRL | Xboundformer · MosMedData+ | Dice: 0.706 -> 0.709 | Helped |
| apply DALE instead of CIRL | Xboundformer · MosMedData+ | 95HD: 22.349 -> 20.534 | Helped |
| add AE&ER soft thresholding (ablation) | DALE on Xboundformer · ISIC2016&ph2 | Dice: 0.9165 -> 0.9359 | Helped |
| add ER to AE-only soft thresholding (ablation) | DALE on Xboundformer · ISIC2016&ph2 | Dice: 0.9198 -> 0.9359 | Helped |
| add AE to ER-only soft thresholding (ablation) | DALE on Xboundformer · ISIC2016&ph2 | Dice: 0.9220 -> 0.9359 | Helped |
| add loss consistency-based confidence parameter ω (ablation) | DALE on Xboundformer · ISIC2016&ph2 | Dice: 0.9225 -> 0.9359 | Helped |
| add unstable representation calibration loss LW (ablation) | DALE on Xboundformer · ISIC2016&ph2 | Dice: 0.9258 -> 0.9359 | Helped |
| add loss consistency-based confidence parameter ω (ablation) | DALE with AE, ER, and LW · ISIC2016&ph2 | Dice: 0.9266 -> 0.9359 | Helped |
| add ER to AE, ω, and LW (ablation) | DALE on Xboundformer · ISIC2016&ph2 | Dice: 0.9276 -> 0.9359 | Helped |
Surprising findings
- BECO, a noisy-label method that replaces fuzzy-region samples with pseudo-labels, underperforms the plain Xboundformer baseline on ISIC2016&ph2 (Dice 0.881 vs 0.917).
Limitations
- Each iteration requires two backpropagation passes, making per-iteration cost high.
- Fuzzy/non-fuzzy partition relies on hyper-parameters (tau, patch size, alpha) that are tuned via sensitivity analysis.
Experimental setup
Datasets: MosMedData+ ISIC-2016 ph2 Kvasir CVC-ClinicDB CVC-ColonDB EndoScene ETIS Synapse
Benchmarks: MosMedData+ ISIC2016&ph2 Polyp (Kvasir, CVC-ClinicDB, CVC-ColonDB, EndoScene, ETIS) Synapse
Metrics: Dice mIoU 95HD ASD
Baselines: BECO GCE MW-Net CIRL U-Net Xboundformer EMCAD LanGuide
Builds on: meta-learning sample reweighting co-teaching CIRL
Closest papers by meaning
- DyCON: Dynamic Uncertainty-aware Consistency and Contrastive Learning for Semi-supervised Medical Image Segmentation (CVPR 2025 · similarity 0.86)
- GapMatch: Bridging Instance and Model Perturbations for Enhanced Semi-Supervised Medical Image Segmentation (AAAI 2025 · similarity 0.84)
- S2S2: Semantic Stacking for Robust Semantic Segmentation in Medical Imaging (AAAI 2025 · similarity 0.84)
- Unified Medical Lesion Segmentation via Self-referring Indicator (CVPR 2025 · similarity 0.84)
- Unleashing the Potential of Vision-Language Pre-Training for 3D Zero-Shot Lesion Segmentation via Mask-Attribute Alignment (ICLR 2025 · similarity 0.83)
- HACDR-Net: Heterogeneous-Aware Convolutional Network for Diabetic Retinopathy Multi-Lesion Segmentation (AAAI 2024 · similarity 0.83)
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.