aiporre / gcn_segmentation

Stroke-GFCN: segmentation of Ischemic brain lesions

Graph-based Fully Convolutional Network (GFCN)

Main script "training.py"

Run from project directory

python traininig.py # runs default MNIST

Datsets

A model is generated in the path ./ with the name {network type}-ds{dataset type}.pth. For example, GFCN-dsGMINST.path.

The dataset options are:

	dataset	contains
1	GVESSEL12	Vessel 12 dataset. CT Scans of 20 patients. Graph grid generated.
2	GMISNT	Segmentation generated from the MNIST dataset. Graph grid generated.
3	GSVESSEL	Vaculature simulations based on the VascuSynth.Graph grid generated.
4	VESSEL12	Vessel 12 dataset. CT Scans of 20 patients
5	MISNT	Segmentation generated from the MNIST dataset
6	SVESSEL	Vaculature simulations based on the VascuSynth.
7	ISLES2018	ISLES challenge dataset. CTP maps and CT scans brain.
8	GISLES2018	ISLES challenge dataset. CTP maps and CT scans brain. Graph grid generated

the available models are:

	model	description
1	GFCN	Graph FCN architecture with pooling and unpooling methods. GFCN16s with barycentric unpooling.
2	GFCNA	Graph FCN architecture with pooling and unspooling methods. GFCN equivalent to the FCN16s with isotropic unpooling
3	GFCNB	Graph FCN architecture with pooling and unspooling methods. GFCN equivalent to the FCN32s isotropic unpooling
4	GFCNC	Graph FCN architecture with pooling and unspooling methods. model G-FCN 8s equivalent isotropic unpooling
5	GFCND	Graph FCN architecture with pooling and unspooling methods. GFCN equivalent to the FCN32s with topk
6	PointNet	PointNet++ Qi et al. (2017)
7	UNet	U-net from Ronneberger et al. (2015)
8	FCN	FCN from Long et al. (2015)
9	DeepVessel	DeepVesselNet from Tetteh et al. (2018)

The available training looses are: BCE, Focal Loss, Generalized Dice Loss.

Evaluation and Training

The evaluation takes the test partition and computes metrics DSC and Acc. Pres. and Recall. It also produces a flat image and a volume with color for TP(green), FP(red) and FN(blue), on the top of the CT scan for the figure. The volume is a raw file with with labeles 1,2,3 for TP(1), FP(2) and FN(3).

To evaluate on a dataset run:

python training.py -s GVESSEL12 -n GFCNC -b 2 -c DCSsigmoid -t True -vd data/vessel12/vessel_g/ -X True

More options

Pre-Transformation The pretransform available is cropping of one lung lobe at between pre_transform = Crop(30,150,256,256), which means from lower-front-left corner (30,150,0) crop a parallelopiped along z-axis with base 256x256

How to set your dataset?

The structure needed is to have a root folder that contains the raw and the processed, in the case of the Graph datasets. in the case of the euclidea, the root folder should just contain the files that correspond to that dataset. The best option is to create a soft link inside the directory data in this repository. For example inside a directory, like for the euclidean case:

./
../
vessel12@ -> /path/to/your/files_mdh

And for the Geometrical you might want to add one more level so:

data
|-gvessel12
  |-./
  |-../
  |- raw@ -> /path/to/your//files_mdh

Cite this work

Ariel Iporre-Rivas, Dorothee Saur, Karl Rohr, Gerik Scheuermann, Christina Gillmann, "Stroke-GFCN: ischemic stroke lesion prediction with a fully convolutional graph network," J. Med. Imag. 10(4) 044502 (17 July 2023) https://doi.org/10.1117/1.JMI.10.4.044502

bibtex

@article{10.1117/1.JMI.10.4.044502,
author = {Ariel Iporre-Rivas and Dorothee Saur and Karl Rohr and Gerik Scheuermann and Christina Gillmann},
title = {{Stroke-GFCN: ischemic stroke lesion prediction with a fully convolutional graph network}},
volume = {10},
journal = {Journal of Medical Imaging},
number = {4},
publisher = {SPIE},
pages = {044502},
keywords = {medical imaging, stroke prediction, machine learning, graph neural networks, multi-modal imaging},
year = {2023},
doi = {10.1117/1.JMI.10.4.044502},
URL = {https://doi.org/10.1117/1.JMI.10.4.044502}
}