crop image in UNETR

[ouyuxuanbridge]

Hi,

I am a little confused about this "allow smaller" in Monai.transforms, does it mean some of the image, especially the margin parts won' t be used in training?

Besides, in my last attempt of UNETR , I tried not to use the random crop, with a image size of 646464 , will the data be loaded in the model still be of the same size?

I deleted these two: (https://github.com/Project-MONAI/tutorials/blob/master/3d_segmentation/unetr_btcv_segmentation_3d.ipynb)
CropForegroundd(keys=["image", "label"], source_key="image"),
RandCropByPosNegLabeld(
keys=["image", "label"],
label_key="label",
spatial_size=(96, 96, 96),
pos=1,
neg=1,
num_samples=4,
image_key="image",
image_threshold=0,
),

Thanks!

I really appreciate your help!

[Nic-Ma]

Hi @Ouyuxuan623 ,

The allow_smaller arg means: when computing box size with margin to crop, whether allow the image size to be smaller than box size, more details: https://github.com/Project-MONAI/MONAI/blob/dev/monai/transforms/croppad/array.py#L706
If True, the cropping can work on even small images.

About the UNETR question, @ahatamiz may help share some comments?

Thanks in advance.

[ahatamiz]

If random crop is not used, then the input will be same as the original image size, whether re-sampled or not, which can be problematic for UNETR, since it expects a fixed size resolution due to its patch embedding layer. I recommend using random crop.