2D Slices from 3D Input

[dannyhow12]

Hello,

Once again, thank you for the amazing repo and the ever so supportive MONAI community.

I have achieved success in implementing the 3D UNet with the MONAI framework, however I would like to now test it on a 2D UNet, where I would wish to input the 3D images that I have, have them sliced into 2D images, and to be trained with MONAI.

I have come across the 2d_inference_3d_volume.ipynb, which gives a suggestion on how to perform so. However, it only uses a dummy example. I would appreciate if any suggestions are given on how I would input my 3D images and incorporate it into the function. Many thanks

Also, by referring to the 2d unet_trainingdict.py, it is able to load the 2D slices, however, how would I want to make sure that the slices that are input to the network are of the same case? Therefore aggregating all the slices of that case, then computing the Dice score

Best regards,
Danny

[Nic-Ma]

Hi @dannyhow12 ,

Thanks for your interest here.
I think the 2d_inference_3d_volume.ipynb tutorial is exactly for your use case.
@surajpaib Could you please help provide more real-world practice?

Thanks in advance.

[dannyhow12]

Hi @wyli @Nic-Ma ,

Thank you so much for the quick updates! I am eternally grateful, I will try it out and revert back if there are any positive or negative advancements regarding this method.

Best regards,
Danny

[dannyhow12]

Hi @wyli @Nic-Ma , hope you all have a great week ahead

I have tried out the section of the code to incorporate it to my dataset. The problem that I face is:

Prerequisite:
My datasets generally are of the size 512 x 512 x z, where z is the varying number of slices in a 3D CT Scan.

1. After applying the first train_transforms, the check_loader returns a shape of [1, 512, 512, 611], which is applicable as the channel is set to be first with AsChannelFirstd, and the last slice is where we want to perform our slicing.

2. At the second data_loader, where it incorporates patch_func and patch_transform. However, when the check_loader returns the shape, it outputs an error of

    File "/usr/local/lib/python3.7/dist-packages/monai/utils/misc.py", line 154, in ensure_tuple_rep
    raise ValueError(f"Sequence must have length {dim}, got {len(tup)}.")
    ValueError: Sequence must have length 2, got 3.
   

By going through the patch_func and patch_transform, it should have already sliced to a 2d patch of 48 by 48. Could it be due to the problem of not defining the axis of my dataset properly?

Best regards,
Danny

[wyli]

you are right, AsChannelFirstd is not robust, I should have used EnsureChannelFirstd. could you please try the same modification as this commit: b4648a5 you should get the output shape for your datasets:

first volume's shape:  torch.Size([1, 1, 512, 512, z]) torch.Size([1, 1, 512, 512, z])
first patch's shape:  torch.Size([3, 1, 48, 48]) torch.Size([3, 1, 48, 48])

[dannyhow12]

Hi @wyli @Nic-Ma ,

1. I do not really understand the 3D transform of RandCropbyPosNegLabeld. 4 samples are generated from each image, with a spatial size of [-1, -1. 1]. How does this achieve slicing all z slices of a specified image? If I understand correctly, it only slices 4 images with a spatial size of 512, 512, 1 (referring to my case where all images are D x H x W of z x 512 x 512). Or is this achieved by the SqueezeDimd? Thank you for your guidance.
2. As each image will be sliced to its corresponding number of slices, thus I would want to use CacheDataset to speed up the training, as in the spleen segmentation notebook example. However, each time when I run the training with CacheDataset caching my image with its transforms, the RAM overwhelms and crashes (Note: I am using Google Colab to train, with a RAM of 12GB). The conventional Dataset works fine, but it is much slower.

Thank you in advance.

Best regards,
Danny

[wyli]

Hi @dannyhow12, the previous tutorial was not trying to address your initial ask. Now I have updated to include another notebook to fetch all z slices which is what you want. However because the number of slices is different across the volumes, we can only use an iterable dataset instead of a regular dataset with known total slices.

https://github.com/Project-MONAI/tutorials/blob/1dfa20ca373a582f404856b3ff509958655b183f/modules/2d_slices_from_3d_training.ipynb

(this notebook requires monai v0.9.0, it'll be officially released early next week.)