FedDBL: Communication and Data Efficient Federated Deep-Broad Learning for Histopathological Tissue Classification

Introduction

The implementation of:

FedDBL: Communication and Data Efficient Federated Deep-Broad Learning for Histopathological Tissue Classification

Abstract

Histopathological tissue classification is a fundamental task in computational pathology. Deep learning-based models have achieved superior performance but centralized training with data centralization suffers from the privacy leakage problem. Federated learning (FL) can safeguard privacy by keeping training samples locally, but existing FL-based frameworks require a large number of well-annotated training samples and numerous rounds of communication which hinder their practicability in the real-world clinical scenario. In this paper, we propose a universal and lightweight federated learning framework, named Federated Deep-Broad Learning (FedDBL), to achieve superior classification performance with limited training samples and only one-round communication. By simply associating a pre-trained deep learning feature extractor, a fast and lightweight broad learning inference system and a classical federated aggregation approach, FedDBL can dramatically reduce data dependency and improve communication efficiency. Five-fold cross-validation demonstrates that FedDBL greatly outperforms the competitors with only one-round communication and limited training samples, while it even achieves comparable performance with the ones under multiple-round communications. Furthermore, due to the lightweight design and one-round communication, FedDBL reduces the communication burden from 4.6GB to only 276.5KB per client using the ResNet-50 backbone at 50-round training. Since no data or deep model sharing across different clients, the privacy issue is well-solved and the model security is guaranteed with no model inversion attack risk.

Requirements

• CUDA
• 1×GPU
• 2×CPU
• Python 3.9.12
• numpy==1.22.4
• pytorch==1.12.0
• torchvision==0.13.0
• scikit-learn==1.1.1

Usage

Installation

• Download the repository.

git clone https://github.com/tianpeng-deng/FedDBL.git

• Install python dependencies.

Data and Backbone Preparation

• We provide an example to train a FedDBL on Multi-center CRC and BCSS, where the dataset can be downloaded from Multi-center CRC and BCSS.
• All backbones (pre-trained models) are stored in load folder. The CTransPath backbone is provided here.
• The architecture of dataset is provided in folder "dataset" and also illustrated below (more details can be seen in /dataset/CRC/partitions.txt):

    |-- dataset
        |-- CRC
        |    |-- 100
        |    |    |-- Fold 1
        |    |    |    |-- Client 1
        |    |    |    |    |-- Train 
        |    |    |    |    |-- Valid
        |    |    |    |-- Client 2
        |    |    |    |-- Client 3
        |    |    |    |-- Client 4
        |    |    |    |-- Centralized
        |    |    |-- Fold 2
        |    |-- 070
        |    |    |-- Fold 1
        |    |    |    |-- Client 1
        |    |    |    |    |-- Train 
        |    |    |    |-- Client 2
        |    |    |    |-- Client 3
        |    |    |    |-- Client 4
        |    |    |    |-- Centralized
        |    |    |-- Fold 2
        |    |-- 050
        |-- BC

There is only one "Valid" for each Fold and every partitions use the same "Valid" for fair comparison.

FedDBL

• To use this example code, you must prepare different proportions of dataset and store them in datset folder. And you can train and test the FedDBL by simply using the command:

python main.py

To run on BCSS, you need to change the weights because there are three users in BCSS but four in Multi-center CRC.

Citation

If you find the code useful, please consider citing our paper using the following BibTeX entry.

@misc{deng2023feddbl,
      title={FedDBL: Communication and Data Efficient Federated Deep-Broad Learning for Histopathological Tissue Classification}, 
      author={Tianpeng Deng and Yanqi Huang and Guoqiang Han and Zhenwei Shi and Jiatai Lin and Qi Dou and Zaiyi Liu and Xiao-jing Guo and C. L. Philip Chen and Chu Han},
      year={2023},
      eprint={2302.12662},
      archivePrefix={arXiv},
      primaryClass={eess.IV}
}