This repository contains the implementation of CHistNet, a deep learning framework designed for histopathology image classification using supervised contrastive learning. This code is associated with the following publication:
"Histopathology Image Classification Using Supervised Contrastive Deep Learning"
Md Mamunur Rahaman, Ewan K. A. Millar, Erik Meijering
University of New South Wales, Sydney, Australia, NSW Health Pathology, St. George Hospital, Kogarah, Australia.
In this study, we introduce CHistNet, a novel deep learning framework that leverages supervised contrastive learning alongside cross-entropy loss to enhance classification performance on histopathology images. The model is first pretrained using contrastive loss to learn robust image representations, followed by fine-tuning with cross-entropy loss for the classification task. This approach achieves state-of-the-art results across multiple histopathology image datasets.
The following datasets were used for training and evaluation:
- BRACS
- BACH
- HE-GHI-DS
- MHIST
- EBHI
Please refer to the paper for detailed instructions on how to download and prepare these datasets.
Before training, the datasets must be converted into TensorFlow Records (TFRecords) to optimize the input pipeline. You can generate the TFRecords by running the generate_tfrecords.py script. Make sure to update the file paths within the script to match your local dataset directories.
- python generate_tfrecords.py
- python main.py
Please cite the following paper:
@inproceedings{rahaman2024histopathology,
title={Histopathology Image Classification Using Supervised Contrastive Deep Learning},
author={Rahaman, Md Mamunur and Millar, Ewan K. A. and Meijering, Erik},
booktitle={IEEE International Symposium on Biomedical Imaging (ISBI)},
year={2024},
organization={IEEE}
}