3D Binary classification using monai, pytorch based framework

[Harris91-lee]

Hello everyone,
My name is Harris. In the past few months i started using Monai framework and i used SEGRESNET model to do single and Multi label segementation. It worked just fine.

However, now im trying to do Binary classification using monai framework as well. I have volume data of good cases in one folder and bad cases in another folder. I also have a txt file that contains the file names and their labels, for ex:
***.nii.gz 0
***.nii.gz 0
***.nii.gz 0
***.nii.gz 1
***.nii.gz 1
...... so on

So my question in short how can i prepare my data and their binary labels?
let's say the volume data will be [1,1,128,128,128], where: 1 referese to number of volume, the other 1 is the channel number, and the 128 is volume dimensions. Should i add the label at the end? so it will be like [1,1,128,128,128,1]? or should i slice the volume to 2D first then do data preparation and model training?
i found same examples for cnn,vgg16,densenet121, ... but mostly they are 2D. So im posting here my question.

I will be grateful to hear some suggestions from you.

Thank you.

[KumoLiu]

Hi @Harris91-lee, I think you can find some 3d classification tutorials in this fold. You can use a dictionary-based transform or array-based transform, it's up to you. The difference is that you should use ImageDataset for the array-based and monai.data.Dataset for the dict-based. You could find more details in the link I posted.
https://github.com/Project-MONAI/tutorials/blob/4d4c565c681fd7df0b01889c9d599b9b73e0b401/3d_classification/torch/densenet_training_dict.py#L33-L60

https://github.com/Project-MONAI/tutorials/blob/4d4c565c681fd7df0b01889c9d599b9b73e0b401/3d_classification/torch/densenet_training_array.py#L32-L57

Hope it can help you, if faced any issues, feel free to ask me, thanks!

[Harris91-lee]

Hello @KumoLiu :)
So im getting good results thanks to you :)
The accuracy is between 92% to 96% depending on the testset.
I wonder if i can improve it to 98% or 99%. Like what will be a good approach to do that?
i already tried batch-size, epoch, resizing 64,96,128, do you have another suggestion? Dataset size interestingly shouldn't big big, i mean now im using dataset with 400 volumes, 200 each.
Many thanks~

[KumoLiu]

Hi @Harris91-lee, I'd like to say that there are always many training tricks to try, but perhaps augmenting the data to increase the number of training sets is a good way to go. Of course, you can also experiment with different loss functions, optimizers, and so on. Thanks!

[Harris91-lee]

Hi @KumoLiu ^^
I also have one more question regarding the network used here for the classification.
Is Densnet121omptimum choice for the MR binary classification? like there are other networks resnet18, vgg16, or even cnn ...etc.
Is there a reason behinde choosing Densnet121 over the other networks?

Thank you in advance ^^

[KumoLiu]

Hi @Harris91-lee, I think there is probably no particular reason to use Densnet121 in the tutorial, just an example :). You could try to use any other networks as your baseline. And from the previous thread, I noticed that you have already had more than 95% accuracy. If your training can accept 100%, you could try to use vgg or any other simpler network. Reducing model complexity can sometimes improve model performance. Thanks.

[Harris91-lee]

I see ~ thank you for your response. I'm getting something between 96~98% now metric accuracy using Densenet121 which is great actually. However, im trying to see if i can get better results. I am willing to try just Densenet without '121' and as you mentioned vgg16.

cheers ^^