vkola-lab / brain2020 Goto Github PK

dear author，I really need your help！I try to use MLP_B in your code on my dataset ,but I can't understand the definition of the demor in the dataloader's MLP_Data. demor = [(demor[0]-70.0)/10.0] + gender + [(demor[2]-27)/2], what is it mean? thanks very much!

in the Data_Preprocess/back_remove.py, Line 55, i think it should be for file in glob(folder + '.npy') , not for file in glob(folder + '.nii'). am i right?
thanks!

This is not exactly an issue, but I couldn't find other means of communicating it. I came across these platform: http://brainchop.org/ and its repo here: https://github.com/neuroneural/brainchop. It might be interesting to know what kind of algorithms runs this.

i have download your code and running it .but i found i got poor accuracy (about 60%)when i run the fcn network and it converenge difficultly.So i curious about what accuracy you got when run the fcn,i don't read about in paper.expect your reply.

Hi Shangran Qiu,
Thank your attention.May I get your help?

1. In "utils.py", "demor = [(demor[0]-70.0)/10.0] + gender + [(demor[2]-27)/2] ". May I ask what is the meaning of this formula and how it was obtained.
2. In "utils.py", "def get_AD_risk(raw)" , " x1, x2 = raw[0, :, :, :], raw[1, :, :, :]". In the data preprocessing part, the .npy file is three-dimensional information, but there is a four-dimensional information, which causes the system to report an error. Therefore, May I ask how to solve it.
   3.In "utils.py", "def get_AD_risk(raw)" , "risk = np.exp(x2) / (np.exp(x1) + np.exp(x2))". May I ask that, why can we get risk from raw data through this equation.
   I appreciate your work very much. Once again, thank you for your time and kindly help. I am looking forward to hearning from you.

Hi, I am using the preproess methods in Data_Preprocess folder to preprocess my dataset,but i don't know the brain_region.ny provided in your code is still work for my dataset( my dataset may not have the same shape as your data)? if not ,could you introduce how to creat a new brain_region.npy file that work for my data?

Thanks for your sharing~ I've met trouble when I followed the script for background removal, it appeared the error as follows:
0%| | 0/2292 [00:00<?, ?it/s]
Traceback (most recent call last):
File "back_remove.py", line 54, in
back_remove(file, temp, '/data/charlen/ADNI/ANDI_remove/')
File "back_remove.py", line 32, in back_remove
and data[new_x, new_y, new_z] < -0.6 and temp[new_x, new_y, new_z] < 0.8:
IndexError: index 179 is out of bounds for axis 0 with size 170

Thank you very much for sharing the source code~
I am currently working on the multi-classification problem of AD, may I ask if it is possible to apply DPMs to the multi-classification situation？

Hi Shangran Qiu
Thank you for providing code for your Brain 2020 paper, but I have the following questions.

1. Random selection 474747 patches can ensure the distinction of features? If AD and NC subjects are very similar in some patches, but giving them different labels during training, which may lead model hard to convergence.
2. In the inference stage, linear layer was transformed into conv layer directly by dense_to_conv(). However, the conv block is local connected structure, the linear layer is a full connection structure. Does it make sense to transform directly?
3. MMSE features directly improved the classification performance of MLP model (according to the formula "DEMOR = [(DEMOR [0]-70.0)/10.0] + GENDER + [(DEMOR [2]-27)/2]", AD_MMSE<0, NC_MMSE>0), but MMSE features have strong prior knowledge, which cannot be obtained directly. Can this model be directly applied in clinical practice?

Hello~ I've followed your work recently, but I'm confused about the MRI data preprocessing according to your code of the 'Data_preprocess'. Can you introduce this part briefly in the readme file? It will be greatly appreciated.

It would be helpful if you can extend the data preprocessing section and include a small point on data directories and where to edit the 'data path' 'them in the provided scripts. ADNI or other data sources tend to have some data organized in a way that doesn't allow the 'Dataloader.py' script to load it directly without some work done on reorganizing the data. This can save time for people conducting the training.

Hi shangranq, I sincerely hope to get your reply. Due to my programming ability, I would like to ask how do I draw the Figure 6. I don't know much about t-SNE, but want to make a visualization of my dataset through t-SNE. Can you provide the code for this，sincerely thanks for your help.

Thank you for providing python scripts for your Brain 2020 paper.
I analyzed ADNI and AIBL data with the scripts and tested, validated, and tested models and obtained results similar to your paper.
I would like to run the scripts only to INFER my data rather than TRAIN and VALIDATE using the already trained FCN and MLP models.
Would you please tell me how to do that?

I need help to run this on my local machine. this repo is pretty disorganized so it is a little hard to follow along. How can I organize this to run it better on my machine?

I am working on a problem using a subset of ADNI data (330 training images) and using your model to train, but I am unable to get good performance on the data:

1: FCN training and validation accuracies always hover around ~50%. The validation confusion matrix indicates that the FCN either classifies all as 0 or 1, and keeps alternating between the two during training.
2: Tried changing learning rate, file_num etc to no use
3: mlp_A which uses FCN and DPMs also produces similar performance (around ~50% over 300 training epochs)
4: mlp_C which uses FCN and clinical features also hovers around that performance
5: I tried to give an easy classification problem to FCN(0s and 1s with linear boundary) and it quickly goes to 100% accuracy which tells me that learning is taking place

Any suggestions you could help me with? Is FCN supposed to do better standalone or does it need more layers? Why do CNN and FCN have 1 as the input number when we are feeding them patches of 47x47?

HI shangran,
Thank your attention. May I ask how can I plot the Fig4A ? Due to my programming ability, I would like to ask how do I draw (iii) based on the riskmap, and how do I superimpose (i) and (ii)? I did not find the code to generate this figure.
Thank you very much

I have a question about the filename in the NACC.csv.

For example, the filename is mri5155ni_s003a1000, what do the s and a represent respectively, is it to select the specified file from multiple nii files in the zip file?

Thanks!

In the registration.py script, line 31, you called a script labeled 'bash registration.sh' However, such a script doesn't exist on this repository. There exists 'registration.sh'

Was any misunderstanding or mistake in this?

The training phase was superb. I would like to ask more about the inference method. Earlier you have mentioned that one can use the 'model wrapper.py' to work with the inference method. I would like to go further and save the model and maybe deploy it. Do you have any suggestions on how I can save the model using the scripts in the 'model wrapper.py'?

By stating that $2 is the filename of the raw data, how is one supposed to be defining this variable? I suppose I can't put one single file name since this script is meant to run in a loop.

Hello, could you please send me the processed or original data? I've been applying here for over a week and haven't come down yet

Preprocessing
step2: convert nifti into ...