Immunofixation Electrophoresis (IFE) Image Recognition based on Deep Learning

This repository provides source code for IFE image recognition in the following paper:

• Honghua Hu, Wei Xu et al. "Expert-level Immunofixation Electrophoresis (IFE) Image Recognition based on Explainable and Generalizable Deep Learning". Clinical Chemistry, 69, no. 2 (2023): 130-139.

BibTeX entry:

@article{Hu2023IFE,
author = {Honghua Hu and Wei Xu and Ting Jiang and Yuheng Cheng and Xiaoyan Tao and Wenna Liu and Meiling Jian and Kang Li and Guotai Wang},
title = {Expert-Level Immunofixation Electrophoresis Image Recognition based on Explainable and Generalizable Deep Learning},
year = {2023},
url = {https://doi.org/10.1093/clinchem/hvac190},
journal = {Clinical Chemistry},
volume = {69},
issue = {2},
pages = {130-139},
}

Requirements

• Pytorch version >=1.9.0
• PyMIC, a Pytorch-based toolkit for medical image computing. Version 0.2.5 is required. Install it by pip install PYMIC==0.2.5.
• Som basic python packages such as Numpy, Pandas, scipy.
• See requirements.txt for more details

Image and preprocess

The images in this study are from two different systems that have different image styles (see data/data_a and data/data_b, respectively). We preprocess the images to make them have the same arrangement and size. The following figures show images before and after preprocessing.

To play with a demo for image preprocessing, run the following command:

python preprocess.py

Demo for inference

To use the pretrained model for inference, download the checkpoints from Google Drive and save them to ckpts. Note that in the oringal paper, each of the three networks (VGG16, ResNet18 and MobileNetv2) has five checkpoints based on 5-fold cross validation. Due to the Google Drive space limit, we only upload one checkpoint for each network for model ensemble. Run the following script for inference:

python demo_inference.py

By defualt, it uses the image data/data_a/20200824_1012358442.jpg from group a as an example. You can set different image names, such as an image from group b by editing line 72:

img_name, group  ="data/data_b/9971568DTouch64.jpg", "b"