The size of the proposed random crop ROI is larger than the image size

[ramchandracheke]

Hi,

Thank you for this wonderful project.

File "/home/ramchand/.local/lib/python3.7/site-packages/monai/transforms/utils.py", line 334, in generate_pos_neg_label_crop_centers
raise ValueError("The size of the proposed random crop ROI is larger than the image size.")

ValueError: The size of the proposed random crop ROI is larger than the image size.
RuntimeError: applying transform <monai.transforms.croppad.dictionary.RandCropByPosNegLabeld object at 0x7ffecef93550>

Could you please help me to solve this error. as My image size is variable(no of slices are different in each nifti file.eg. (342, 342, 91) and I am using Unet. (128,128,128). I am using following combination of transform.

train_transforms = Compose(
[
LoadImaged(keys=["image", "label"]),
AddChanneld(keys=["image", "label"]),
Spacingd(keys=["image", "label"], pixdim=(
1.5, 1.5, 1.0), mode=("bilinear", "nearest")),
Orientationd(keys=["image", "label"], axcodes="RAS"),
ScaleIntensityRanged(
keys=["image"], a_min=-400, a_max=1500,
b_min=0.0, b_max=1.0, clip=True,
),
CropForegroundd(keys=["image", "label"], source_key="image"),
RandCropByPosNegLabeld(
keys=["image", "label"],
label_key="label",
spatial_size=(128,128,128),
pos=1,
neg=1,
num_samples=4,
image_key="image",
image_threshold=0,
),ToTensord(keys=["image", "label"]),
]
)

Thanks a million!

[rijobro]

Hi @ramchandracheke, your example image had 91 slices in the z-dimension. This is already smaller than the 128 that you've requested of RandCropByPosNegLabeld. You're also using Spacing and CropForegroundd before calling RandCropByPosNegLabeld, so it's hard to know what size your images are before RandCropByPosNegLabeld.

To debug this, why don't you comment out the final RandCropByPosNegLabeld transform. You can then run through all of your data and check the size. Something like (completely untested):

min_sizes = (10000, 10000, 10000)
for d in dataset:
    shape = d["image"].shape[1:]
    for i in range(len(shape)):
        if shape[i] < min_sizes[i]:
            min_sizes[i] = shape[i]
print(min_sizes)

[ramchandracheke]

Hi @rijobro,
Yeah. Thank you for your reply. My data-set consist of different volumes with different number of slices. Therefore, I found that the volume with minimum dimension of (512,512,74) to (512,512,974). Yes you spotted right that, the dimensions of the image is reduced further after using CropForegroundd. Therefore, I am just using below transform.

train_transforms = Compose([
        LoadImaged(keys=["image", "label"]),
        AddChanneld(keys=["image", "label"]),
        ScaleIntensityRanged(
            keys=["image"], a_min=-400, a_max=1500,
            b_min=0.0, b_max=1.0, clip=True,
        ),
        RandSpatialCropd(keys=["image", "label"], roi_size=(128, 128,128), random_size=False),
        SpatialPadd(keys=["image", "label"],spatial_size=(128,128,128), mode ='reflect'),

        ToTensord(keys=["image", "label"]),
    ]
)

This is working, however I think using RandCropByPosNegLabeld I could get better segmentation.

Thank you very much!

[rijobro]

If you want to use RandCropByPosNegLabeld, could you do SpatialPadd first, to make sure that everything is at least (128, 128, 128), and then use RandCropByPosNegLabeld?

By the way, I think in your current set up, calling SpatialPadd after RandSpatialCropd is redundant, since your images will already be (128, 128, 128) after RandSpatialCropd, so SpatialPadd won't do anything.

[ramchandracheke]

Hi @rijobro,

Thanks for your inputs. I changed the order to --> CropForegroundd, -->RandCropByPosNegLabeld and --> Rand3DElasticd and it works for me. I observed that when I added Rand3DElasticd the loss is almost constant loss where as its decreasing steadily without using using Rand3DElasticd augmentation. In segmentation task its elastic deformation is used widely as a data augmentation. However, in this case the network converging very slow. Do you have any suggestion on this?

Thanks a million!

[rijobro]

Would be a good idea to visualise your images after the elastic transformation. If you get the parameters wrong, the resulting image can look very different compared to the original.

[ramchandracheke]

Yeah... That's true!