Awesome

HATNet: Self-attention-based end-to-end network for diagnosing histopathological images

This repository contains the source code of our work for classifying histopathological images using hollistic attention network (HATNet). Breifly, we introduce an attention-based network to classify breast biopsy images. We streamline the histopathological image classification pipeline and show how to learn representations from gigapixel size images end-to-end. Our network extends the bag-of-words approach and uses self-attention to encode global information, allowing it to learn representations from clinically relevant tissue structures without any explicit supervision.

Table of contents

1. Visualization
2. Dependencies
3. Training and Evaluation
4. Pre-trained models
5. Citation
6. Issues

Visualization

For the breast cancer diagnosis, several clinical studies have shown that ductal regions and stromal tissues are an importnat biomarkers. Our analysis reveals that our network pays attention to clinically relevant structures. See more details in the paper.

Ducts

Below figure shows that our network is able to identify ducts of variable size and texture as an important structure. Here, ductal regions (marked by pathologists) are shown in red while top-50% bags predicted by our network are shown in blue.

Stoma

Below figure shows that our network is able to identify stroma as an important tissue. Here, each sub-figure is organized from left to right as: breast biopsy image, stroma tissue (shown in pink) labeled by pathologists, and top-50% words (words that belong to stroma tissue are shown in pink while the remaining words are shown in blue) identified using our model. The remaining 50% words are shown in white.

Dependencies

Our source code requires:

• OpenSlide (for reading whole slide images)
• PyTorch (for deep learning)
• Other python libraries like scikit, opencv.

Note: We have tested our code in Python3 only. If you are using Python2, please make appropriate changes.

Install OpenSlide

You can install OpenSlide using below commands:

sudo apt-get install openslide-tools
sudo apt-get install python-openslide
pip install openslide-python
 

We have tested our code only on linux environments. If you are facing issues installing OpenSlide, please check here:

https://openslide.org/

Install PyTorch

Please see below webpage for installing PyTorch. We have tested our code with PyTorch 1.0+.

https://pytorch.org/

Install other dependencies

Other dependencies can be installed by running the below command:

pip install -r requirements.txt 

Training and Evaluation

See scripts training_script.sh and eval_script.sh for training and evaluation related examples. Our evaluation code evaluates the performance of a model across different metrics, including plots for ROC curves. Please see cmat_metrics.py and evaluator.py files.

Note: For averaging model checkpoints, see average_checkpoints.py file.

Pre-trained models

Check ./model_zoo directory for pre-trained models.

Citation

Please cite as:

@misc{mehta2020hatnet,
    title={HATNet: An End-to-End Holistic Attention Network for Diagnosis of Breast Biopsy Images},
    author={Sachin Mehta and Ximing Lu and Donald Weaver and Joann G. Elmore and Hannaneh Hajishirzi and Linda Shapiro},
    year={2020},
    eprint={2007.13007},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}

Issues

Thanks for your interest in our work. For any issues, please raise a request.