QFedPolyp: A Communication- and Inference-Efficient Federated Learning Framework for Polyp Segmentation
QFedPolyp introduces a federated learning framework for polyp segmentation that combines quantization-aware training with low-precision communication to reduce bandwidth costs. The approach uses lightweight U-Nets trained locally at hospitals, transmitting only 8-bit quantized parameters to a central server for aggregation via Federated Averaging. Quantized models achieve up to 4x reduction in transmission cost and 1.5x faster inference while maintaining competitive Dice scores (0.910 on Kvasir-
Analysis
TL;DR
- QFedPolyp introduces a federated learning framework for polyp segmentation that combines quantization-aware training with low-precision communication to reduce bandwidth costs.
- The approach uses lightweight U-Nets trained locally at hospitals, transmitting only 8-bit quantized parameters to a central server for aggregation via Federated Averaging.
- Quantized models achieve up to 4x reduction in transmission cost and 1.5x faster inference while maintaining competitive Dice scores (0.910 on Kvasir-SEG, 0.930 on CVC-ClinicVideoDB).
- The method enables privacy-preserving collaboration across institutions without sharing sensitive medical data, supporting real-time clinical deployment.
Why It Matters
This work addresses critical bottlenecks in deploying AI for healthcare: data privacy, communication overhead, and inference latency. By enabling efficient federated learning with quantization, it makes collaborative model training feasible across hospitals with limited bandwidth and strict compliance requirements—directly impacting the scalability of AI-driven diagnostic tools in real-world clinical settings.
Technical Details
- Framework integrates quantization-aware training into local U-Net optimization, simulating low-precision weights during training to minimize accuracy loss upon quantization.
- Clients transmit 8-bit uniform quantized model parameters instead of full-precision floats, reducing communication volume by ~75% per round.
- Central server reconstructs quantized weights using dequantization before applying Federated Averaging to update global model.
- Evaluated on four public polyp segmentation datasets: Kvasir-SEG, CVC-ClinicVideoDB, PolypGen, and BKAI-IGH NeoPolyp.
- Lightweight U-Net architecture optimized for edge-device inference, balancing accuracy and computational efficiency.
Industry Insight
Healthcare AI providers should adopt quantized federated frameworks like QFedPolyp to overcome regulatory and infrastructural barriers in multi-institutional collaborations. The demonstrated speedup and bandwidth savings make this approach viable for resource-constrained clinics, accelerating adoption of computer-aided diagnosis systems while preserving patient data sovereignty. Future implementations may explore adaptive bit-width quantization or dynamic client selection to further optimize performance.
Disclaimer: The above content is generated by AI and is for reference only.