3D Classification Task

[Meddebma]

Hi,

I am trying to make a 3D classification Task with Images that have a mask and a label. Therefore I used the ImageDataset passing the images, segmentation masks and labels in it. My Idea is that the classification should not be done for the whole image, but only the part of the image that has been selected (In this case it is the spleen). For a better understanding I have two diseases of the spleen in CT scans, and I would like to make a classifier to differentiate between them.

Here is my Notebook:
https://github.com/Meddebma/pyradiomics/blob/master/3D_Classification.ipynb

my Questions:

1. When I use the ImageDataset, should I use a different transform for the segmentation mask?
2. How should the original image be cropped so that only the spleen is gonna be isolated for the classification?
3. Here is the Error I get during training:
   Expected object of scalar type Long but got scalar type Float for argument NifTI image reader 2D/3D #2 'target' in call to _thnn_nll_loss_forward

Thank you

[Nic-Ma]

Hi @Meddebma ,

I thnk maybe you can use MaskIntensityd transform to mask the spleen region:
https://github.com/Project-MONAI/MONAI/blob/dev/monai/transforms/intensity/dictionary.py#L743
Maybe @rijobro @ericspod can share more research approaches to train the model.

Thanks.

[Nic-Ma]

Could you please help share the error log?

Thanks.

[Meddebma]

There was unfortunately no error log, after circa 10min the RAM crashed and the session restarted and the cell output showed:
Loading dataset: 0%| | 0/2 [00:00<?, ?it/s]
is it maybe because I used CacheDataset?

Thanks

[Nic-Ma]

Hi @Meddebma ,

As I said in the previous thread: the returned data of partition_dataset is 2 items: train_data and val_data, so I think your CacheDataset can set as:

datasets = partition_dataset_classes(data, classes, shuffle=False, ratios=[2, 1])
train_ds = CacheDataset(data=datasets[0], ...)
val_ds = CacheDataset(data=datasets[1], ...)

Thanks.

[Meddebma]

Hi @Nic-Ma,

it worked perfectly, and even the accuracy went up to 0.96 during the training !!!

Thank you so much for your effort and help!

[Nic-Ma]

No problem, glad to see your great progress!
Please feel free to raise a ticket if you face any other questions.

Thanks.