This project is protected under the EULA form you will sign during the registration from ADP Website to access the database. Use of this repository code or any biproduct (model weights, pre-trained architecture, etc) is strictly prohibited for any commercial use under the EULA agreement.
This repository contains the Keras implementation of Convolutional Neural Network (CNN) training on the Atlas of Digital Pathology (ADP) database. For more information and how to access the ADP database, please visit Atlas Project.
Patch-level Annotated Database of Digital Pathology images for Histological Tissue Type (HTT) Classification, as presented by Hosseini et al.'s CVPR 2019 paper Atlas of Digital Pathology: A Generalized Hierarchical Histological Tissue Type-Annotated Database for Deep Learning.
- HTTs with no training examples (i.e. N.G.A, N.G.O, N.G.E, N.G.R, N.G.T) are removed, to prevent infinite class weights
- (Optional) Log-inv-freq used instead of Inv-freq as class weights
- More models are included for training, including a tailored specific model "HistoNet" where the architecure is simplified for the purpose of training for ADP
- Additional option for Colour Agumentation is provided using either HSV or YCbCr method
These instructions will get you a copy of the project up and running on your local machine for development and testing purposes. You will download a copy of the ADP database and train/test all CNN architectures with any of the five taxonomic configurations.
Mandatory
python(checked on 3.5)pandas(checked on 0.23.4)keras(checked on 2.2.4)numpy(checked on 1.16.2)tensorflow(checked on 1.13.1)
Optional
matplotlib(checked on 3.0.2)
cd folder/to/clone-into
git clone https://github.com/mahdihosseini/ADP.git
First, download the separate ADP database to a local directory. To access the database, please refer to ADP Website
Then, open demo_train.py, then edit the value of DATASET_DIR to the location of that local directory.
Edit the settings.csv file as necessary. This file allows you to specify the following
- Model type/Variant: defines the CNN architecture for training. Available options are: VGG/Default, ResNet/(resnet_18 or resnet_34, or resnet_50), Xception/Xception_V1, MobileNet/V1, Inception/Default, HistoNet/Series-1.0
- Level: level of training using an of five taxonomic configurations. Available options are: L1, L2, L2+, L3, L3+
- Dataset type: the ground-truth labels set to: ADP-Release1-Flat
- Micron Resolution: resolution of images for training/validation set to: 1 (The ADP is released with 1 micron resolution)
- Downsampling Method: images are downsampled from 0.25 micron resolution to 1 micron using bicubic and set to: bicubic
- CLR: cyclical learning rate. Avaialble options are: TRUE/FALSE
- Colour Augmentation: using colour augmentation for training. Available options are: TRUE/FALSE (by default we set this to YCbCr augmentation. You can substitute HSV in learner.py line 211 to 'custom_augmentation_ycbcr' for ImageDataGenerator)
- Epoch: maximum numbe of epochs
Next, run the demo script to train
cd folder/to/clone-into
python demo_train.py
@inproceedings{hosseini2019atlas,
title={Atlas of digital pathology: A generalized hierarchical histological tissue type-annotated database for deep learning},
author={Hosseini, Mahdi S and Chan, Lyndon and Tse, Gabriel and Tang, Michael and Deng, Jun and Norouzi, Sajad and Rowsell, Corwyn and Plataniotis, Konstantinos N and Damaskinos, Savvas},
booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
pages={11747--11756},
year={2019}
}
- @geifmany for his implementation of VGG16 for CIFAR, which we adapted for our VGG16 implementation