Research map / Label-efficient and robust learning / Domain adaptation and generalization
FAMNet: Frequency-aware Matching Network for Cross-domain Few-shot Medical Image 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 FAMNet, a frequency-aware prototype matching network for cross-domain few-shot medical image segmentation, evaluated on cross-modality, cross-sequence (cardiac) and cross-institution-style settings.
Problem
Few-shot medical image segmentation models fail under domain shift from different imaging techniques (e.g., CT vs MRI, LGE vs b-SSFP), where support-query bias and inter-domain differences degrade segmentation.
Approach
Decompose support and query foreground features into low, mid and high frequency bands via FFT, apply distinct attention weightings (similar-component emphasis for domain-agnostic bands, suppression for domain-specific bands), then fuse bands with cross-attention guided by the mid-frequency band to suppress domain-variant information.
Key results
- On cross-modality CT to MRI, FAMNet reaches 65.79% mean Dice, versus 63.01% for the second-best method (SSL-ALP).
- On cross-modality MRI to CT, mean Dice is 64.75%, versus 57.29% for PATNet.
- On cross-sequence, mean Dice reaches 71.58% (LGE to b-SSFP) and 61.39% (b-SSFP to LGE), surpassing the second-best by 10.14% and 1.33%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FAMNet rather than SSL-ALP | Abdominal CT → MRI · Abdominal CT → MRI | Dice: 63.01 -> 65.79 | Helped |
| Use FAMNet rather than PATNet | Abdominal MRI → CT · Abdominal MRI → CT | Dice: 57.29 -> 64.75 | Helped |
| Use FAMNet rather than IFA | Cardiac LGE → b-SSFP · Cardiac LGE → b-SSFP | Dice: 62.28 -> 71.58 | Helped |
| Use FAMNet rather than PATNet | Cardiac b-SSFP → LGE · Cardiac b-SSFP → LGE | Dice: 60.06 -> 61.39 | Helped |
| Add CPG to PANet baseline (ablation) | PANet baseline · Abdominal CT → MRI | Dice: 32.46 -> 53.97 | Helped |
| Add FAM to the CPG-equipped baseline (ablation) | PANet baseline with CPG · Abdominal CT → MRI | Dice: 53.97 -> 62.02 | Helped |
| Add MSF to the CPG- and FAM-equipped baseline (ablation) | PANet baseline with CPG and FAM · Abdominal CT → MRI | Dice: 62.02 -> 65.79 | Helped |
| Use attention rather than inverse attention for the high-frequency band (ablation) | FAMNet · Abdominal CT → MRI | Dice: 65.79 -> 61.29 | Hurt |
| Use attention rather than inverse attention for the low-frequency band (ablation) | FAMNet · Abdominal CT → MRI | Dice: 65.79 -> 60.02 | Hurt |
| Use attention rather than inverse attention for both low- and high-frequency bands (ablation) | FAMNet · Abdominal CT → MRI | Dice: 65.79 -> 58.68 | Hurt |
Surprising findings
- Applying positive attention weighting to any domain-specific frequency band decreased performance, indicating over-reliance on source-domain support-query correlation harms generalization.
Limitations
- Evaluated only on three cross-domain settings with abdominal and cardiac data; cross-institution results deferred to supplementary material.
- Pseudo-masks from supervoxel clustering are used for supervision rather than manual labels in meta-learning.
Experimental setup
Datasets: Abdominal MRI (20 T2-SPIR scans) Abdominal CT (20 scans) Cardiac LGE/b-SSFP MRI (45 + 45 scans) Prostate T2-weighted MRI (UCLH, NCI) for cross-institution
Benchmarks: Abdominal CT to MRI and MRI to CT (cross-modality) Cardiac LGE to b-SSFP and b-SSFP to LGE (cross-sequence)
Metrics: Dice score
Baselines: PANet SSL-ALPNet ADNet QNet CATNet RPT PATNet IFA
Builds on: PANet PATNet SSL-ALPNet CATNet ADNet
Closest papers by meaning
- Lightweight Frequency Masker for Cross-Domain Few-Shot Semantic Segmentation (NeurIPS 2024 · similarity 0.88)
- Dual-Agent Optimization framework for Cross-Domain Few-Shot Segmentation (CVPR 2025 · similarity 0.87)
- Adapt Before Comparison: A New Perspective on Cross-Domain Few-Shot Segmentation (CVPR 2024 · similarity 0.86)
- MedFlowSeg: Flow Matching for Medical Image Segmentation with Frequency-Aware Attention (NeurIPS 2026 · similarity 0.85)
- Pancakes: Consistent Multi-Protocol Image Segmentation Across Biomedical Domains (NeurIPS 2025 · similarity 0.85)
- Constructing and Exploring Intermediate Domains in Mixed Domain Semi-supervised Medical Image Segmentation (CVPR 2024 · 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.