
...

NAKO Dataset：  
- Over 3400 labeled MRI images from thousands patients
- Over 500 MRI images for evaluation  

<img src="https://github.com/JiaPeng1234/MRI-Segmentation-Transformer/blob/master/imgs/overview.png?raw=true" width="30%">  

Used network architectures including 3d u-net, non-local neural network, attention u-net are proposed.  

Arxiv:  

...

It can capture the long-range dependencies between pixels more properly, check the paper from Wang Xiaolong https://arxiv.org/abs/1711.07971  

Its architecture as following:  

<img src="https://github.com/JiaPeng1234/MRI-Segmentation-Transformer/blob/master/imgs/non-local.PNG?raw=true" width="70%">

### 3d U-net as baseline
Baseline architecture is a 4-stages 3d u-net, as following:  

<img src="https://github.com/JiaPeng1234/MRI-Segmentation-Transformer/blob/master/imgs/u-net.PNG?raw=true" width="70%">  

## Results
Achieve an average accurancy of 97% of all classes.  

<img src="https://github.com/JiaPeng1234/MRI-Segmentation-Transformer/blob/master/imgs/results1.png?raw=true" width="60%">  

<img src="https://github.com/JiaPeng1234/MRI-Segmentation-Transformer/blob/master/imgs/exp001shape.PNG?raw=true" width="50%"> 

	a/README.md		b/README.md
	...		...
3		3
4	NAKO Dataset：	4	NAKO Dataset：
5	- Over 3400 labeled MRI images from thousands patients	5	- Over 3400 labeled MRI images from thousands patients
6	- Over 500 MRI images for evaluation	6	- Over 500 MRI images for evaluation
7		7
8	<img src="imgs/overview.png" width="30%">	8	<img src="https://github.com/JiaPeng1234/MRI-Segmentation-Transformer/blob/master/imgs/overview.png?raw=true" width="30%">
9		9
10	Used network architectures including 3d u-net, non-local neural network, attention u-net are proposed.	10	Used network architectures including 3d u-net, non-local neural network, attention u-net are proposed.
11		11
12	Arxiv:	12	Arxiv:
13		13
	...		...
37		37
38	It can capture the long-range dependencies between pixels more properly, check the paper from Wang Xiaolong https://arxiv.org/abs/1711.07971	38	It can capture the long-range dependencies between pixels more properly, check the paper from Wang Xiaolong https://arxiv.org/abs/1711.07971
39		39
40	Its architecture as following:	40	Its architecture as following:
41		41
42	<img src="imgs/non-local.PNG" width="70%">	42	<img src="https://github.com/JiaPeng1234/MRI-Segmentation-Transformer/blob/master/imgs/non-local.PNG?raw=true" width="70%">
43		43
44	### 3d U-net as baseline	44	### 3d U-net as baseline
45	Baseline architecture is a 4-stages 3d u-net, as following:	45	Baseline architecture is a 4-stages 3d u-net, as following:
46		46
47	<img src="imgs/u-net.PNG" width="70%">	47	<img src="https://github.com/JiaPeng1234/MRI-Segmentation-Transformer/blob/master/imgs/u-net.PNG?raw=true" width="70%">
48		48
49	## Results	49	## Results
50	Achieve an average accurancy of 97% of all classes.	50	Achieve an average accurancy of 97% of all classes.
51		51
52	<img src="imgs/results1.png" width="60%">	52	<img src="https://github.com/JiaPeng1234/MRI-Segmentation-Transformer/blob/master/imgs/results1.png?raw=true" width="60%">
53		53
54	<img src="imgs/exp001shape.PNG" width="50%">	54	<img src="https://github.com/JiaPeng1234/MRI-Segmentation-Transformer/blob/master/imgs/exp001shape.PNG?raw=true" width="50%">