Communication- and Inference-Efficient Federated Learning for Polyp Segmentation

Madan Baduwal, Priyanka Paudel, Tilak Neupane
Department of Computer Science and Engineering, Mississippi State University
Paper Code Datasets Citation
Problem solution overview comparing centralized training with the proposed federated learning framework.

Hospitals keep colonoscopy images private, train locally with quantization-aware learning, and exchange quantized model updates instead of raw data.

Abstract

Automatic polyp segmentation in colonoscopy images plays an important role in computer-aided diagnosis and early detection of colorectal cancer. Most deep learning approaches rely on centralized training, which requires sharing medical data across hospitals and raises privacy concerns. Federated learning enables collaborative model training without transferring raw patient data, but repeated transmission of high-precision model parameters introduces significant communication overhead.

This work proposes a communication- and inference-efficient federated learning framework for collaborative polyp segmentation. Each hospital locally trains a lightweight U-Net using quantization-aware training and sends quantized model updates to a central server for FedAvg aggregation. Across Kvasir-SEG, CVC-ClinicVideoDB, PolypGen, and BKAI-IGH NeoPolyp, uniform 8-bit communication reduces payload by about 4x while maintaining competitive segmentation accuracy, and the INT8 model delivers up to 1.5x faster inference.

Main Contributions

  • Privacy-preserving federated learning for collaborative polyp segmentation across hospitals.
  • Quantization-aware local training for robust low-precision parameter exchange.
  • Communication-efficient federated updates with strong performance under 8-bit and mixed-precision settings.
  • Deployment-ready INT8 model with faster inference on both CPU and GPU.

Federated Architecture

The server distributes a global segmentation model to participating hospitals, each client trains locally on its private data, then transmits quantized updates back for aggregation.

Federated learning architecture for polyp segmentation.

Lightweight U-Net

A compact encoder-decoder U-Net backbone is used to balance segmentation performance, communication cost, and efficient deployment.

Small U-Net architecture used for polyp segmentation.

Selected Results

Full-precision federated learning reaches Dice scores of 0.910 on Kvasir-SEG and 0.930 on CVC-ClinicVideoDB. Uniform 8-bit communication reduces transmission cost by approximately 4x while maintaining competitive performance at 0.895 and 0.915 Dice on the same datasets. Mixed-precision settings further show that hospitals can operate under different communication budgets with only limited accuracy loss.

Configuration Avg. Bits Comm. Reduction Kvasir Dice CVC Dice PolypGen Dice BKAI Dice
Uniform 8-bit 8.0 4x 0.895 0.915 0.825 0.885
Mixed 8/8/8/16/16 11.2 2.9x 0.899 0.920 0.830 0.894
Uniform 32-bit 32.0 1x 0.910 0.930 0.845 0.905

32-bit Federated Learning

Dice curve for 32-bit federated learning.

32-bit IoU

IoU curve for 32-bit federated learning.

8-bit Federated Learning

Dice curve for 8-bit federated learning.

8-bit IoU

IoU curve for 8-bit federated learning.

Inference Efficiency

Precision CPU Latency CPU FPS GPU Latency GPU FPS Speedup
FP32 77.20 ms 12.95 3.76 ms 266.2 1.0x
INT8 71.59 ms 13.97 2.50 ms 400.0 1.5x

Datasets

The framework is evaluated on four public colonoscopy benchmarks covering multiple acquisition conditions and clinical environments.

Citation

@misc{baduwal2026federatedpolyp,
  title  = {Communication- and Inference-Efficient Federated Learning for Polyp Segmentation},
  author = {Baduwal, Madan and Paudel, Priyanka and Neupane, Tilak},
  year   = {2026},
  note   = {Project page and manuscript draft},
  url    = {https://github.com/madanbaduwal/Communication-Efficient-federated-Learning-For-Polyp-Segmentation}
}