Research map / Continual learning / Domain and prompt-based
FR²Seg: Continual Segmentation Across Multiple Sites via Fourier Style Replay and Adaptive Consistency Regularization
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 FR2Seg, a continual multi-site medical segmentation method that replays old-site styles via Fourier amplitude swapping and regularizes domain-invariant parameters with a transferability-weighted Fourier consistency term.
Problem
Continual medical image segmentation across sequentially arriving sites suffers catastrophic forgetting of previous sites and poor generalization to new sites, while privacy and storage limit rehearsal of raw data.
Approach
Uses low-frequency Fourier amplitude as a domain-style representation: stores averaged old-site amplitude styles to synthesize pseudo images with old-site styles (Fourier Style Replay), and identifies domain-invariant parameters via a Fourier style mixup consistency loss, weighted by cross-site transferability measured by cumulative Dice scores.
Key results
- On Prostate in-order, FR2Seg reaches AVG 88.30% and BWT -1.43%, versus 85.14% AVG for the strongest baseline S3R.
- On Fundus in-order, FR2Seg reaches AVG 85.18% and BWT -1.11%, versus 84.08% for S3R.
- Ablation on Prostate in-order: removing FSR drops AVG to 85.45%, removing FACR to 81.61%, and removing ATW to 87.01%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Fourier Style Replay (FSR) (ablation) | FR2Seg · Prostate and Fundus, in-order and in-reverse | AVG: 85.45 -> 88.30; 75.49 -> 79.68; 84.75 -> 85.18; 84.67 -> 85.38 | Helped |
| add Fourier Style Replay (FSR) (ablation) | FR2Seg · Prostate and Fundus, in-order and in-reverse | BWT: -4.43 -> -1.43; -8.47 -> -3.01; -2.43 -> -1.11; -3.59 -> -3.11 | Helped |
| add Fourier Adaptive Consistency Regularization (FACR) (ablation) | FR2Seg · Prostate and Fundus, in-order and in-reverse | AVG: 81.61 -> 88.30; 74.84 -> 79.68; 84.14 -> 85.18; 85.10 -> 85.38 | Helped |
| add Fourier Adaptive Consistency Regularization (FACR) (ablation) | FR2Seg · Prostate and Fundus, in-order and in-reverse | BWT: -10.59 -> -1.43; -14.63 -> -3.01; -4.01 -> -1.11; -4.77 -> -3.11 | Helped |
| add Adaptive Inter-Domain Transferability Weighting (ATW) (ablation) | FR2Seg · Prostate and Fundus, in-order and in-reverse | AVG: 87.01 -> 88.30; 78.67 -> 79.68; 84.73 -> 85.18; 85.21 -> 85.38 | Helped |
| add Adaptive Inter-Domain Transferability Weighting (ATW) (ablation) | FR2Seg · Prostate and Fundus, in-order and in-reverse | BWT: -1.98 -> -1.43; -5.51 -> -3.01; -1.70 -> -1.11; -4.37 -> -3.11 | Helped |
| retain all six previous models instead of only the latest model (ablation) | FR2Seg · Prostate, in-order | AVG: 86.74 -> 88.30 | Helped |
| retain all six previous models instead of only the latest model (ablation) | FR2Seg · Prostate, in-reverse | AVG: 78.85 -> 79.68 | Helped |
Limitations
- Storing previous-site models adds storage overhead that grows with the number of sites
- Evaluated only on two public 2D segmentation benchmarks
Experimental setup
Datasets: Prostate MRI (NCI-ISBI13, I2CVB, PROMISE12) Fundus (REFUGE, RIM-ONE-r3, Drishti-GS)
Benchmarks: Prostate MRI multi-site sequence Fundus multi-site sequence
Metrics: AVG (seen-domain average Dice) BWT (backward transfer) Dice score
Baselines: FineTune JointTrain LWF POD PLOP MASLR EWC RWalk nnUNet-ViT S3R
Builds on: CISR nnUNet EWC Learning without Forgetting
Closest papers by meaning
- MedCRP-CL: Continual Medical Image Segmentation via Bayesian Nonparametric Semantic Modality Discovery (ICML 2026 · similarity 0.86)
- Beyond Background Shift: Rethinking Instance Replay in Continual Semantic Segmentation (CVPR 2025 · similarity 0.86)
- Representation Compensation Networks for Continual Semantic Segmentation (CVPR 2022 · similarity 0.85)
- SimCS: Simulation for Domain Incremental Online Continual Segmentation (AAAI 2024 · similarity 0.84)
- Knowledge Transfer Scaling Laws for 3D Medical Imaging (NeurIPS 2026 · similarity 0.84)
- LangDAug: Langevin Data Augmentation for Multi-Source Domain Generalization in Medical Image Segmentation (ICML 2025 · 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.