QFedPolyp: Federated Learning Framework Cuts Polyp Segmentation Costs 4x

Centralized deep learning for medical image segmentation requires hospitals to share sensitive patient data, raising serious privacy concerns. Federated learning solves this by keeping data local, but it introduces huge communication overhead from repeatedly exchanging full-precision model parameters. QFedPolyp, developed by Madan Baduwal and Priyanka Paudel, tackles this trade-off head-on. The framework combines quantization-aware training with low-precision communication: each hospital trains a lightweight U-Net on its private data while simulating quantization during training. Instead of sending 32-bit floats, clients transmit 8-bit quantized parameters to a central server, where they are reconstructed and aggregated using Federated Averaging. This preserves the privacy benefits of federated learning while dramatically reducing bandwidth.

Tested on four polyp segmentation datasets (Kvasir-SEG, CVC-ClinicVideoDB, PolypGen, and BKAI-IGH NeoPolyp), QFedPolyp achieved Dice scores of 0.910 on Kvasir-SEG and 0.930 on CVC-ClinicVideoDB—competitive with full-precision federated training. The uniform 8-bit communication reduced transmission cost by approximately 4 times, and the quantized models ran up to 1.5 times faster during inference. This means hospitals can collaborate on training a robust segmentation model without sharing raw images, using less network bandwidth and achieving faster predictions for real-time clinical workflows. The lightweight models are well-suited for deployment in resource-constrained settings, such as endoscopy suites or mobile diagnostics.