Model doesn't learn

[gkrisp98]

Hi, I am new to the field of medical images. I am working with 435 MRIs and I have to classify them to 5 classes. I have created a subject dataset both for nifti and DICOM files but neither of them is learning. The training and validation accuracy range from 20 to 22% which is just random.
I used the transforms RandomAffine(), RescaleIntensity(), RandomNoise and CropOrPad() to make the shape of the objects (160,224,224) as it was (160,384,384).
As of now I used a efficient net 3D, a resnet3d and a resnet18 from monai but none of them seems to learn anything.
Is there any advice for what I can do to make my models start learning ?

[Nic-Ma]

Hi @ericspod @dongyang0122 ,

I think you guys are deep learning research experts, could you please help share some advice here?

Thanks in advance.

[dongyang0122]

Hi @gkrisp98,

It could be potentially multiple reasons causing an ineffective learning process. First, you can check the step for data pre-processing (since you did not mention in your descriptions). MRI has greater appearance variation compared to CT images. And the transform "NormalizeIntensityd" is recommended to use for MRI as data pre-processing. Second, you can simply the problem to make sure the learning process is proper. For instance, you can convert the 5-class recognition problems into a 2-class one (via merging some class labels), or disable the data augmentation at the moment, to make sure that the model is learning something. Last not the least, you can further check your scripts and compare to MONAI examples (https://github.com/Project-MONAI/tutorials/tree/master/3d_classification). It is possible that some necessary lines are missing to cause such ineffective learning.

BTW, if you can share your scripts, we can take a quick look and provide some insights.

Best,
Dong

[gkrisp98]

Ok I will run some epochs without these lines and I will post the curves. Thanks a lot for your advices!

[gkrisp98]

So, here are the curves with early stopping at 10 epochs

:

[dongyang0122]

It now appears that the training is starting to take effect. The training loss starts to drop and the accuracy rises.

[gkrisp98]

Then should I disable early stopping and just run it for 50 (or more) epochs ? But shoudn't the val loss and acc still be saturating ?

[dongyang0122]

You can try to disable early stopping and see if the model overfits the data.