ViFlT-assisted Histopathology: From H&E Style Standardization to Virtual Fiber Image Transformation

🔥🔥ViFIT Paper Published in Medical Image Analysis (MIA).

Authors

Shu Wang#, Xiao Zhang#, Xingfu Wang#, Chenyong Lv, Xiahui Han, Xiong Lin, Deyong Kang, Ruolan Lin, Liwen Hu, Feng Huang*, Wenxi Liu*, Jianxin Chen*

Pytorch implementation of ViFIT

OS Requirements

• Linux: Ubuntu 18.04
• Python 3.7 + Pytorch 1.8.1
• NVIDIA GPU + CUDA 11.0 CuDNN 8

Installation Guide

Install PyTorch 0.4+ and other dependencies (e.g., torchvision, visdom and dominate).

• For pip users, please type the command: pip install -r requirements.txt.
• For Conda users, please create a new Conda environment using conda env create -f environment.yml.

Download Guide

Download the required datasets and weights:

• Public multi-style H&E datasets: BreaKhis, BACH, BreCaHAD
• H&E-to-fiber test dataset: H&E-to-fiber dataset
• Model weights used in this study: weights

Data Preconditioning

Before training, split the large images into 256×256 pixel patches. For Stage1 with the BreaKHis dataset, modify the patch script to assign labels for multi-style images. Run the following command to preprocess the data:

python patch.py

Organize the data into the following directory structure:

• Training data for SourceA: ./datasets/data/trainA
• Training data for SourceB: ./datasets/data/trainB
• Test data for SourceA: ./datasets/data/testA
• Test data for SourceB: ./datasets/data/testB
• Test results for SourceA: ./datasets/data/testC

There are two main stages in our framework

• Stage1: Style Standardization
• Stage2: Fiber Transformation To switch between stages:
• Navigate to Stage1:
  • cd Stage1
• Navigate to Stage2:
  • cd Stage2

Stage1 Training and Testing

Train the model for style standardization:

python train.py --dataroot ./datasets/data --name name --cls --lab --input_nc 4 --output_nc 4

Test the model and save results:

python test.py --dataroot ./datasets/data --name name --cls --hebing --chutu ./datasets/data/testC --input_nc 4 --output_nc 4 --lab --shuchu 'fake_B'

Stage2 Training and Testing

Train the model for fiber transformation:

python train.py --dataroot ./datasets/data --name name --Auxiliary 1

Test the model and save results:

python test.py --dataroot ./datasets/data --name name --cls --hebing --chutu ./datasets/data/testC --shuchu 'fake_B'

Combined ViFIT Testing

To run the full pipeline (Stage1 + Stage2), use the provided script:

sh ViFIT.sh

Intermediate results are saved in the ./checkpoints/ directory, and the .pth files will be stored in their respective folders.

Contact

If you have any questions, please contact Shu Wang at shu@fzu.edu.cn or Wenxi Liu at wenxiliu@hotmail.com.