Research map / Federated learning and privacy / Personalized and heterogeneous FL
Dual-calibrated Co-training Framework for Personalized Federated Semi-Supervised 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 PFSSL, a personalized federated semi-supervised framework for medical image segmentation that aggregates per-client personalized models via a rule relationship graph and co-trains them with a private local model using local (boundary patch replacement) and global (gradient-direction) calibration.
Problem
Federated semi-supervised medical image segmentation suffers from client data heterogeneity and from semi-supervised methods failing to adapt across clients, causing knowledge bias in local models and unstable convergence.
Approach
Combine personalized federated aggregation (a rule relationship graph weighting model similarity, dataset size, and model quality) with a single-teacher-dual-student co-training scheme, plus two calibration modules: boundary-patch pseudo-label replacement locally and cosine-similarity-based gradient calibration globally.
Key results
- On LiverSeg23 the method reaches 84.16% DC and 85.00% JC, improving DC by 12.69 points over HSSF.
- On Polyp the method reaches 73.86% DC, versus 71.72% for the FedAvg upper bound with full labels.
- Local calibration alone raises DC on LiverSeg23 from 81.21 to 87.20 in the fusion comparison (Table 3) relative to its baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PFSSL with personalized aggregation, dual-student co-training, and dual calibration instead of FedIRM | U-Net with ResNet-34 · LiverSeg23 | DC: 82.74 -> 84.16 | Helped |
| Use PFSSL with personalized aggregation, dual-student co-training, and dual calibration instead of FedAvg+MT | U-Net with ResNet-34 · ISIC2018 | DC: 86.35 -> 89.44 | Helped |
| Use PFSSL with personalized aggregation, dual-student co-training, and dual calibration instead of FedAvg (upper-bound) | U-Net with ResNet-34 · Polyp | DC: 71.72 -> 73.86 | Helped |
| Add local calibration (ablation) | PFSSL on LiverSeg23 · LiverSeg23 | DC: 77.39 -> 80.76 | Helped |
| Add personalized aggregation (ablation) | PFSSL on LiverSeg23 · LiverSeg23 | DC: 77.39 -> 81.77 | Helped |
| Add global calibration (ablation) | PFSSL on LiverSeg23 · LiverSeg23 | DC: 77.39 -> 77.81 | No clear effect |
| Add local calibration and personalized aggregation (ablation) | PFSSL on LiverSeg23 · LiverSeg23 | DC: 77.39 -> 83.48 | Helped |
| Add local calibration and global calibration (ablation) | PFSSL on LiverSeg23 · LiverSeg23 | DC: 77.39 -> 81.61 | Helped |
| Add personalized aggregation and global calibration without local calibration (ablation) | PFSSL on LiverSeg23 · LiverSeg23 | DC: 77.39 -> 74.15 | Hurt |
| Add all three components: local calibration, personalized aggregation, and global calibration (ablation) | PFSSL on LiverSeg23 · LiverSeg23 | DC: 77.39 -> 84.16 | Helped |
| Replace local calibration fusion with MiDSS fusion (ablation) | Semi-supervised LiverSeg23 model · LiverSeg23 | DC: 81.21 -> 83.73(+2.52) | Helped |
| Replace local calibration fusion with UCMT fusion (ablation) | Semi-supervised LiverSeg23 model · LiverSeg23 | DC: 81.21 -> 85.08(+3.87) | Helped |
Surprising findings
- Applying global calibration on top of personalized aggregation decreased performance, which the authors attribute to local model instability.
- Naive combinations of personalized FL and semi-supervised methods performed worse than some simpler baselines, and on Polyp most baselines fell near the lower bound due to data drift.
Limitations
- Relies on a private proprietary dataset (LiverSeg23) that is not released, limiting reproducibility.
- Hyperparameters (rho, mu, eta) tuned on specific datasets, with global calibration unstable when combined with personalized aggregation.
Experimental setup
Datasets: LiverSeg23 ISIC-2018 Polyp (Jha et al. 2020; Bernal et al. 2015)
Benchmarks: LiverSeg23 ISIC-2018 Polyp
Metrics: Dice coefficient Jaccard index HD95 ASSD
Baselines: FedIRM HSSF PgFed+dhc FedU+bcp FedAvg+MT FedProx+FixMatch MiDSS UCMT
Builds on: Mean Teacher Cross Pseudo Supervision FedAvg FedU PgFed
Closest papers by meaning
- FedVCK: Non-IID Robust and Communication-Efficient Federated Learning via Valuable Condensed Knowledge for Medical Image Analysis (AAAI 2025 · similarity 0.87)
- (FL)$^2$: Overcoming Few Labels in Federated Semi-Supervised Learning (NeurIPS 2024 · similarity 0.86)
- Overcoming Heterogeneous Data in Federated Medical Vision-Language Pre-training: A Triple-Embedding Model Selector Approach (AAAI 2025 · similarity 0.86)
- FedBiP: Heterogeneous One-Shot Federated Learning with Personalized Latent Diffusion Models (CVPR 2025 · similarity 0.86)
- FedA3I: Annotation Quality-Aware Aggregation for Federated Medical Image Segmentation against Heterogeneous Annotation Noise (AAAI 2024 · similarity 0.86)
- Federated Modality-Specific Encoders and Multimodal Anchors for Personalized Brain Tumor Segmentation (AAAI 2024 · similarity 0.86)
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.