ehosseiniasl / 3d-convolutional-network Goto Github PK

Alzheimer disease diagnosis by Deeply Supervised 3D Convolutional Network

Python 100.00%

3d-convolutional-network alzheimer-disease-diagnostics deep-learning

3d-convolutional-network's Introduction

Alzheimer Disease Diagnosis by Deeply Supervised 3D Convolutional Network

Diagnosing Alzheimer disease from 3D MRI T1 scans from ADNI dataset. The initial results using 3D Convolutional Network is published in ICIP 2016 [1]. The second model used deeply supervision to boost the performance on all binary and three-way classification of AD/MCI/Normal classes. The results are published on arxiv [2]

Using Transfer Learning

Pretraining 3D CNN with 3D Convolutional Autoencoder on source domain
Finetuning uper fully-connected layers of 3D CNN using supervised fine-tuning on target domain
Using deeply supervision in supervised fine-tuning of upper fully-connected layers

DATA

List of all subject ids are in ADNI_subject_id directory

###Papers

[1] E. Hosseini-Asl, R. Keynton and A. El-Baz, "Alzheimer's disease diagnostics by adaptation of 3D convolutional network," 2016 IEEE International Conference on Image Processing (ICIP), Phoenix, AZ, USA, 2016, pp. 126-130.
[2] E. Hosseini-Asl, G. Gimel'farb, and A. El-Baz, “Alzheimer's Disease Diagnostics by a Deeply Supervised Adaptable 3D Convolutional Network”, arXiv:1607.00556 [cs.LG, q-bio.NC, stat.ML], 2016.

3d-convolutional-network's People

Contributors

Stargazers

Watchers

3d-convolutional-network's Issues

Hi how to see fc layer's weights?

After I finetuned my scae, I got my scae_AD_Normal.pkl, but I find that only 2 elements in there(I only use one layer CAE1+2 layers fc+softmax), the size of it are (8,3,1,3,3) and (8,), when I use load_fc() function, it goes wrong, when did the fc weights save?

Can you show me how to train your code?

I want to use it to classify AD and Normal, but I have some trouble of how to train your code, can you tell me how to do that?

Load mat file

Dear Author
I'm working on analysing your paper code(3D-CNN for Alzheimer's disease diagnosis).
But I'm facing some problems on loading mat file.
In main.py, in 814 line, you load mat file.
I'm trying to transform MRI data on mat file, but I can't achieve that, due to lack of understanding about the structure of mat file that you are using in your code.

In short,

How you are transforming Nifti file to mat file?
What's the structure of the mat file?

I can't train my model!

When I pre_trained my model,
First I define a stacked_CAE3d，then I load CAE, my batch_size=4, and I set my labels are 0,1,0,1
the prob is always equals to 0.5, and it never change!
I don't know why!

Thank you for your code!

Thank you so much for your code! But I am confusing to understand how to use your code, can you give me an example about the code? Much appreciated.

Training of CAE

I am trying to implement your paper in Torch and wanted some help in it.

Firstly I want to know how are you exactly training the CAE, greedy layerwise or end to end.

when you are decoding, you use convolution, but why the weights are not flipped over?

when you are encoding, use the code:
self.hidden_layer=ConvolutionLayer3D(rng,
input=self.inputs,
signal_shape=signal_shape,
filter_shape=filter_shape,
act=activation,
border_mode='full',
if_hidden_pool=False)

when you are decoding, use the code:
self.recon_layer=ConvolutionLayer3D(rng,
input=self.hidden_layer.output,
signal_shape=self.hidden_image_shape,
filter_shape=self.hidden_filter_shape,
act=activation,
border_mode='valid')
Your paper write To reduce the number of the free parameters, the decoding, P k , and encoding, W k , weights were tied by flipping over all their dimensions as proposed in.
So I want to ask, why there is no flipping over in the code? Or did I miss something?

Hi I want to know what is the cost when you doing your pre_training CAE?

When I doing my pre_training CAE2, my cost is about 100 every data, I think it is abnormal, what is your cost when you doing your pre_train?

How to run

Could you please tell me how to run this code?

Using my own data

Hello, can anyone please help me with running this project with my own data?

Thank you much!

when I use do_pretraining_cae function,I just get the cost are nan?

why the cost is nan and it never change?

Run time error

(virtual-py2) faisal@DeepLearning-PC:~/Downloads/ts$ python main.py
Traceback (most recent call last):
File "main.py", line 22, in
from convnet_3d import CAE3d, stacked_CAE3d
File "/home/faisal/Downloads/ts/convnet_3d.py", line 14, in
import maxpool3d
File "/home/faisal/Downloads/ts/maxpool3d.py", line 7, in
from theano.tensor.signal.downsample import DownsampleFactorMax
ImportError: No module named downsample
Please give me correct solution. Thanks.