Research map / Detection and segmentation / CNN features
Modality-agnostic Domain Generalizable Medical Image Segmentation by Multi-Frequency in Multi-Scale Attention
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 MADGNet, a medical image segmentation network with a multi-frequency multi-scale attention block (MFMSA) and an ensemble sub-decoding module (E-SDM) for deep supervision, evaluated for modality-agnostic and domain-generalizable segmentation across 15 datasets.
Problem
Medical image segmentation models lack generalization across imaging modalities and unseen clinical domains, overlook frequency variance, and suffer information loss from drastic upsampling in multi-task learning with deep supervision.
Approach
Combine multi-frequency channel attention (2D DCT statistics) with multi-scale spatial attention in a single block, and train region, boundary and distance-map tasks in an ensemble cascade (forward and backward streams) rather than parallel heads to compensate for upsampling information loss.
Key results
- On seen-setting binary segmentation across five modalities, MADGNet improves over M2SNet by 1.1% DSC and 1.0% mIoU on average.
- On unseen clinical settings, MADGNet gains 1.9% DSC and 2.7% mIoU over FRCUNet on STU.
- E-SDM ensemble decoding improves unseen-colonoscopy DSC from 73.7 (parallel with deep supervision) to 80.9 (full ensemble with backward stream).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| combine multi-frequency and multi-scale features rather than using single-scale or single-frequency variants (ablation) | MADGNet · CVC-ClinicDB and Kvasir | DSC: 90.7 (0.4), 92.3 (0.5), 92.3 (0.3) -> 93.9 (0.6), 90.7 (0.8) | Helped |
| use multi-frequency rather than single-frequency features at single scale (ablation) | MADGNet · CVC-ClinicDB and CVC-300 | DSC: 90.7 (0.4) -> 92.3 (0.5); 81.7 (0.3) -> 83.3 (1.4) | Helped |
| use multi-scale rather than single-scale features at single frequency (ablation) | MADGNet · CVC-ClinicDB and CVC-300 | DSC: 90.7 (0.4) -> 92.3 (0.3); 81.7 (0.3) -> 84.2 (1.2) | Helped |
| use ensemble rather than parallel deep-supervision predictions (ablation) | MADGNet · CVC-ClinicDB and Kvasir (seen) | DSC: 90.8 -> 91.4 | Helped |
| use ensemble rather than parallel deep-supervision predictions (ablation) | MADGNet · CVC-300, CVC-ColonDB and ETIS (unseen) | DSC: 73.7 -> 80.9 | Helped |
| add a backward stream to ensemble decoding (ablation) | MADGNet · CVC-ClinicDB and Kvasir (seen) | DSC: 91.4 -> 92.0 | Helped |
| add a backward stream to ensemble decoding (ablation) | MADGNet · CVC-300, CVC-ColonDB and ETIS (unseen) | DSC: 77.5 -> 80.9 | Helped |
| add ensemble deep supervision instead of using region prediction only (ablation) | MADGNet · CVC-300, CVC-ColonDB and ETIS (unseen) | DSC: 75.2 -> 80.9 | Helped |
| replace the parallel task-prediction flow with ensemble decoding (ablation) | MADGNet · CVC-300, CVC-ColonDB and ETIS (unseen) | DSC: 76.2 -> 80.9 | Helped |
Surprising findings
- Parallel multi-task deep supervision degraded performance relative to no deep supervision on unseen colonoscopy datasets (DSC 73.7 vs 75.2).
Limitations
- Authors note memory efficiency needs improvement for real clinical use.
- Gains over strongest baselines are modest on some unseen datasets (e.g., BUSI vs FRCUNet, CVC-300 vs M2SNet).
Experimental setup
Datasets: ISIC2018 COVID19-1 BUSI DSB2018 CVC-ClinicDB Kvasir REFUGE PH2 COVID19-2 STU MonuSeg2018 CVC-300
Benchmarks: ISIC2018 COVID19-1 BUSI DSB2018 CVC-ClinicDB Kvasir PH2 COVID19-2 STU MonuSeg2018 CVC-300 CVC-ColonDB
Metrics: DSC mIoU p-value (one-tailed t-test)
Baselines: UNet Attention UNet UNet++ CENet TransUNet FRCUNet MSRFNet HiFormer DCSAUNet M2SNet
Builds on: FCANet UNet++ M2SNet ResNeSt
Closest papers by meaning
- Learning Generalized Medical Image Segmentation from Decoupled Feature Queries (AAAI 2024 · similarity 0.86)
- UniMRSeg: Unified Modality-Relax Segmentation via Hierarchical Self-Supervised Compensation (NeurIPS 2025 · similarity 0.85)
- Modality-Agnostic Structural Image Representation Learning for Deformable Multi-Modality Medical Image Registration (CVPR 2024 · similarity 0.85)
- Domain Generalized Medical Landmark Detection via Robust Boundary-Aware Pre-Training (AAAI 2025 · similarity 0.85)
- Pancakes: Consistent Multi-Protocol Image Segmentation Across Biomedical Domains (NeurIPS 2025 · similarity 0.85)
- Why Invariance is Not Enough for Biomedical Domain Generalization and How to Fix It (NeurIPS 2026 · 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.