Research map / Detection and segmentation / CNN features
ConDSeg: A General Medical Image Segmentation Framework via Contrast-Driven Feature Enhancement
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 ConDSeg, a two-stage medical image segmentation framework (ResNet-50 encoder) with consistency-reinforced encoder pretraining, semantic information decoupling into fg/bg/uncertainty, contrast-driven feature aggregation, and size-aware decoders, reporting SOTA on five datasets.
Problem
Medical image segmentation suffers from ambiguous "soft" boundaries and poor illumination/contrast, and from co-occurrence patterns that mislead models into predicting spurious multiple entities.
Approach
Combines a strong-augmentation consistency pretraining of the encoder (with a binarized BCE consistency loss instead of KL/JS), decouples encoder features into foreground, background and uncertainty branches with complementarity loss, uses foreground/background contrast to guide windowed attention fusion across levels, and assigns decoders to entity size scales to avoid co-occurrence errors.
Key results
- On Kvasir-Sessile, ConDSeg reaches 81.2 mIoU and 89.1 mDSC, versus 76.4 mIoU / 84.2 mDSC for the next best method (DTAN).
- On ISIC-2017, ConDSeg reports 80.9 mIoU and 88.3 mDSC, versus 77.1 / 85.5 for EIU-Net.
- Ablation on GlaS: adding SID+CDFA raises mIoU from 78.3 to 83.3 and SA-Decoder raises it further to 85.1.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ConDSeg instead of the strongest listed baseline, DTAN | Medical image segmentation · Kvasir-Sessile | mDSC: 84.2 -> 89.1 | Helped |
| Use ConDSeg instead of the strongest listed baseline, DTAN | Medical image segmentation · Kvasir-SEG | mDSC: 90.4 -> 90.5 | Helped |
| Use ConDSeg instead of the strongest listed baseline, DTAN | Medical image segmentation · GlaS | mDSC: 87.9 -> 91.6 | Helped |
| Use ConDSeg instead of EIU-Net | Medical image segmentation · ISIC-2016 | mDSC: 91.9 -> 92.5 | Helped |
| Use ConDSeg instead of EIU-Net | Medical image segmentation · ISIC-2017 | mDSC: 85.5 -> 88.3 | Helped |
| Add CR to Stage 1 Net_0 training (ablation) | Net_0 · Kvasir-SEG | mDSC: 85.1 -> 86.1 | Helped |
| Use two-stage training instead of Stage 2 only (ablation) | ConDSeg · Kvasir-SEG | mDSC: 88.9 -> 89.2 | Helped |
| Add CR to two-stage training (ablation) | ConDSeg · Kvasir-SEG | mDSC: 89.2 -> 90.5 | Helped |
| Add SID and CDFA to the baseline (ablation) | ConDSeg baseline · GlaS | mDSC: 87.5 -> 90.4 | Helped |
| Add SA-Decoder to SID and CDFA (ablation) | ConDSeg · GlaS | mDSC: 90.4 -> 91.6 | Helped |
| Increase Stage 1 training from 50 ep. to 100 ep. (ablation) | ConDSeg · Kvasir-SEG | mDSC: 90.6 -> 90.5 | No clear effect |
| Vary CDFA window size; K=3 is best among tested sizes (ablation) | ConDSeg · Kvasir-Sessile | mDSC: 88.4; 89.1; 86.0; 85.9 | Mixed |
Limitations
- Reported gains are evaluated only on five datasets at a fixed 256x256 resolution with simple augmentation.
- Ablations for the main module comparisons are run on a single dataset (GlaS or Kvasir-SEG).
Experimental setup
Datasets: Kvasir-SEG Kvasir-Sessile GlaS ISIC-2016 ISIC-2017
Benchmarks: Kvasir-SEG Kvasir-Sessile GlaS ISIC-2016 ISIC-2017
Metrics: mIoU mDSC recall precision
Baselines: U-Net U-Net++ Attn U-Net PraNet TGANet DCSAU-Net XBoundFormer CASF-Net DTAN CENet CPFNet FAT-Net
Builds on: U-Net PraNet TGANet Grad-CAM
Closest papers by meaning
- DyCON: Dynamic Uncertainty-aware Consistency and Contrastive Learning for Semi-supervised Medical Image Segmentation (CVPR 2025 · similarity 0.88)
- S2S2: Semantic Stacking for Robust Semantic Segmentation in Medical Imaging (AAAI 2025 · similarity 0.87)
- Interactive Medical Image Segmentation: A Benchmark Dataset and Baseline (CVPR 2025 · similarity 0.87)
- MAPSeg: Unified Unsupervised Domain Adaptation for Heterogeneous Medical Image Segmentation Based on 3D Masked Autoencoding and Pseudo-Labeling (CVPR 2024 · similarity 0.86)
- Training Like a Medical Resident: Context-Prior Learning Toward Universal Medical Image Segmentation (CVPR 2024 · similarity 0.86)
- Rethinking U-Net: Task-Adaptive Mixture of Skip Connections for Enhanced Medical Image Segmentation (AAAI 2025 · similarity 0.85)
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.