Error adapting Spleen example to different shaped dataset

[project-monai]

Error adapting Spleen example to different shaped dataset

[ericspod]

I suspect your network configuration isn't compatible with patches that are that small. You're creating your UNet with

model = UNet(
    dimensions=3,
    in_channels=1,
    out_channels=2,
    channels=(16, 32, 64, 128, 256),
    strides=(2, 2, 2, 2),
    num_res_units=2,
    norm=Norm.BATCH,
).to(device)

With this configuration your network will have 5 layers and with strides of 2 for each downsampling you will halve the spatial dimensions of the input image 4 times. This means your image sizes in the network are 20**3 -> 10**3 -> 5**3 -> 2**3 -> 1**3. Attempting to upsample this in the decode path of the network will not work because of the default padding in the upsample convolutions.

I would suggest using patch sizes that are multiples of powers of 2 which, for a dimension of M*2**N allows you to downsample N times. In your case what you can do is use a patch size of 32 and stack your volumes only once so the depth dimension is 40, or double each slice in your volume to get the same.

Alternatively you can stick with 20**3 as your patch size and use (64,128,256) as your channels argument and (2,2) as your strides argument to make a shallower network and see how that works.

[ericspod]

For Spacing you provide a multiple which reduces the size so if you want to go from 1200x340x20 to 1200x340x40 you want something like this:

from monai.transforms import Spacing
s=Spacing(pixdim=(1, 1, 0.49))
t=torch.rand(1,1200,340,20)
print(s(t)[0].shape)  # (1, 1200, 340, 40)

For using Spacingd you would have the same pixdim values as this example.

I'm honestly not sure what to comment on your particular problem, I shall ask my colleagues to see if anyone has something specific to this particular segmentation problem to contribute.

[Irme]

Hello!
I work with sparse data (cerebral microbleed segmentation), and I would recommend you look into the following:

• sampling techniques: oversampling the minority class. This is an easy way of balancing the dataset, but may increase false positive predictions or overfitting.
• data augmentation: in particular in combination with a good sampling technique, this can be a very powerful tool. Data augmentation (flipping, rotating, etc) can have the same effect as oversampling the minority class, but it should also reduce the risk of overfitting.
• loss function: because of the imbalance, I would recommend a loss function designed for imbalanced data, such as generalised dice, tversky loss or focal loss.

Generally with very imbalanced data it takes a lot of patience to fine-tune the network.

[catskillsresearch]

@Irme thanks. Please consider joining the CTF NF Hackathon which is still open and has plenty of medical data. Here are some details:

https://nfhack-platform.bemyapp.com/#/event
We have more than 500 participants registered from around the world as well as more than 40 mentors and 21 projects in progress. Registration remains open throughout the Hackathon and projects can be submitted up until November 13th. 

[catskillsresearch]

Thanks very much Eric. There are 20 planes in the MRI. Should I be trying to "thicken" the plane volumes so that they correspond more to the physical dimensions of the X-Y plane, or is that irrelevant? If so, how do I do that? My MRI images are 20 NumPy bitmaps. I think something like this is done in the Spleen tutorial notebook here:

Spacingd(keys=["image", "label"], pixdim=(1.5, 1.5, 2.0), mode=("bilinear", "nearest")),

The Spleen example from NVidia uses a Nifti format that comes with metadata including an Affine transform that the Spacingd transform accesses. What would be the corresponding arguments for Spacing transform?

Also, one last question. The tumors (in this case, from Neurofibromatosis 1) are very sparse in the whole MRI volume. So 98% of the volume is non-tumor and 2% is tumor. Do you have any thoughts on how to increase accuracy for this very sparse class? I have only 50 full labelled examples to work with. Should I explore Waterloo's "Less than one shot" learning technique?

[catskillsresearch]

Thanks Eric! Intuitively I don't think it should matter. The spleen tutorial has the pixdim=(1.5, 1.5, 2.0). I'm curious what motivated that. It seems like they were doing it to reduce the size of the inputs:

from monai.transforms import Spacing
import torch
s=Spacing(pixdim=(1.5, 1.5, 2.0))
t=torch.rand(1,226,257,113)
print(s(t)[0].shape)  # (1, 151, 172, 57)

Regarding the overall number of samples (50 in my case), and sparsity in the image set of positive labels (tiny tumors), any thoughts will be very welcome.

This is in the context of the Children's Tumor Foundation Hack for NF which is still open for participation.