A 3D multi-modal medical image segmentation library in PyTorch

We strongly believe in open and reproducible deep learning research. Our goal is to implement an open-source medical image segmentation library of state of the art 3D deep neural networks in PyTorch. We also implemented a bunch of data loaders of the most common medical image datasets. This project started as an MSc Thesis and is currently under further development. Although this work was initially focused on 3D multi-modal brain MRI segmentation we are slowly adding more architectures and data-loaders.

Top priorities 21-07

[Update] 21-07 We have just received a brand new GPU. The project developedment was postponed due to lack of computational resources. We will be back with more updates. Please Watch our Github repository for releases to be notified. We are always looking for passionate open-source contributos. Full credits will be given.

• Project restructure, API/CLI design ++
• Minimal test prediction example with pre-trained models
• Overlapping and non-overlapping inference
• Finalize preprocessing on Brats datasets
• Save produced 3d-total-segmentation as nifty files
• Medical image decathlon dataloaders
• StructSeg 2019 challenge dataloaders
• More options for 2D architectures
• Rewrite manual
• New notebooks with google colab support

Quick Start

• If you want to quickly understand the fundamental concepts for deep learning in medical imaging, we strongly advice to check our blog post. We provide a general high-level overview of all the aspects of medical image segmentation and deep learning. For a broader overview on MRI applications find my latest review article.

• To grasp more fundamental medical imaging concepts, check out our post on coordinate systems and DICOM images.

• For a more holistic approach on Deep Learning in MRI you may advice my thesis this.

• Alternatively, you can create a virtual environment and install the requirements. Check the installation folder for more instructions.

• You can also take a quick glance at the manual.

• If you do not have a capable environment or device to run this projects then you could give Google Colab a try. It allows you to run the project using a GPU device, free of charge. You may try our Colab demo using this notebook:

Implemented architectures

• U-Net3D Learning Dense Volumetric Segmentation from Sparse Annotation Learning Dense Volumetric Segmentation from Sparse Annotation

• V-net Fully Convolutional Neural Networks for Volumetric Medical Image Segmentation

• ResNet3D-VAE 3D MRI brain tumor segmentation using auto-encoder regularization

• U-Net Convolutional Networks for Biomedical Image Segmentation

• SkipDesneNet3D 3D Densely Convolutional Networks for Volumetric Segmentation

• HyperDense-Net A hyper-densely connected CNN for multi-modal image segmentation

• multi-stream Densenet3D A hyper-densely connected CNN for multi-modal image segmentation

• DenseVoxelNet Automatic 3D Cardiovascular MR Segmentation with Densely-Connected Volumetric ConvNets

• MED3D Transfer learning for 3D medical image analysis

• HighResNet3D On the Compactness, Efficiency, and Representation of 3D Convolutional Networks: Brain Parcellation as a Pretext Task

Implemented medical imaging data-loaders

Task	Data Info/ Modalities	Train/Test	Volume size	Classes	Dataset size (GB)
Iseg 2017	T1, T2	10 / 10	144x192x256	4	0.72
Iseg 2019	T1, T2	10 / 13	144x192x256	4	0.75
MICCAI BraTs2018	FLAIR, T1w, T1gd,T2w	285 / -	240x240x155	9 or 4	2.4
MICCAI BraTs2019	FLAIR, T1w, T1gd,T2w	335 / 125	240x240x155	9 or 4	4
Mrbrains 2018	FLAIR, T1w, T1gd,T2w	8	240x240x48	9 or 4	0.5
IXI brain development Dataset	T1,T2 no labels	581	(110~150)x256x256	-	8.7
MICCAI Gleason 2019 Challenge	2D pathology images	~250	5K x 5K	-	2.5

Preliminary results

Visual results on Iseg-2017

Iseg and Mr-brains

Model	# Params (M)	MACS(G)	Iseg 2017 DSC (%)	Mr-brains 4 classes DSC (%)
Unet3D	17 M	0.9	93.84	88.61
Vnet	45 M	12	87.21	84.09
DenseNet3D	3 M	5.1	81.65	79.85
SkipDenseNet3D	1.5 M	31	-	-
DenseVoxelNet	1.8 M	8	-	-
HyperDenseNet	10.4 M	5.8	-	-

Usage

How to train your model

• For Iseg-2017 :

python ./examples/train_iseg2017_new.py --args

• For MR brains 2018 (4 classes)

python ./examples/train_mrbrains_4_classes.py --args

• For MR brains 2018 (8 classes)

python ./examples/train_mrbrains_9_classes.py --args

• For MICCAI 2019 Gleason Challenge

python ./examples/test_miccai_2019.py --args

• The arguments that you can modify are extensively listed in the manual.

Inference

How to test your trained model in a medical image

python ./tests/inference.py --args

Covid-19 segmentation and classification

We provide some implementations around Covid-19 for humanitarian purposes. In detail:

Classification model

• COVID-Net A Tailored Deep Convolutional Neural Network Design for Detection of COVID-19 Cases from Chest Radiography Images

Datasets

Classification from 2D images:

• COVID-CT dataset

• COVIDx dataset

3D COVID-19 segmentation dataset

• COVID-19 CT Lung and Infection Segmentation Dataset

Latest features (06/2020)

• On the fly 3D total volume visualization
• Tensorboard and PyTorch 1.4+ support to track training progress
• Code cleanup and packages creation
• Offline sub-volume generation
• Add Hyperdensenet, 3DResnet-VAE, DenseVoxelNet
• Fix mrbrains,Brats2018,Brats2019, Iseg2019, IXI,MICCAI 2019 gleason challenge dataloaders
• Add confusion matrix support for understanding training dynamics
• Some Visualizations

Support

If you really like this repository and find it useful, please consider (★) starring it, so that it can reach a broader audience of like-minded people. It would be highly appreciated :) !

Contributing to Medical ZOO

If you find a bug, create a GitHub issue, or even better, submit a pull request. Similarly, if you have questions, simply post them as GitHub issues. More info on the contribute directory.

Current team

Ilias Papastatis, Sergios Karagianakos and Nikolas Adaloglou

License , citation and acknowledgements

Please advice the LICENSE.md file. For usage of third party libraries and repositories please advise the respective distributed terms. It would be nice to cite the original models and datasets. If you want, you can also cite this work as:

@MastersThesis{adaloglou2019MRIsegmentation,
author = {Adaloglou Nikolaos},
title={Deep learning in medical image analysis: a comparative analysis of
multi-modal brain-MRI segmentation with 3D deep neural networks},
school = {University of Patras},
note="\url{https://github.com/black0017/MedicalZooPytorch}",
year = {2019},
organization={Nemertes}}

Acknowledgements

In general, in the open source community recognizing third party utilities increases the credibility of your software. In deep learning, academics tend to skip acknowledging third party repos for some reason. In essence, we used whatever resource we needed to make this project self-complete, that was nicely written. However, modifications were performed to match the project structure and requirements. Here is the list of the top-based works: HyperDenseNet model. Most of the segmentation losses from here. 3D-SkipDenseNet model from here. 3D-ResNet base model from here. Abstract model class from MimiCry project. Trainer and Writer class from PyTorch template. Covid-19 implementation based on our previous work from here. MICCAI 2019 Gleason challenge data-loaders based on our previous work from here. Basic 2D Unet implementation from here.Vnet model from here