Why the data shape are different through transform "spacingd.

[wyli]

Discussed in #5934

^{Originally posted by xxsxxsxxs666 February 3, 2023}

val_transforms = Compose(
    [
        LoadImaged(keys=["image", "label"]),
        EnsureChannelFirstd(keys=["image", "label"]),
        ScaleIntensityRanged(
            keys=["image"], a_min=-57, a_max=400,
            b_min=0.0, b_max=1.0, clip=True,
        ),
        CropForegroundd(keys=["image", "label"], source_key="image"),  # default: value>0
        Orientationd(keys=["image", "label"], axcodes="RAI"),
        Spacingd(keys=["image", "label"], pixdim=(
            0.35, 0.35, 0.5), mode=("bilinear", "nearest")),  # Now only Sequential_str, How to add spline
    ]
)

The shape and metadata are the same between image and label

this problem only happen when I use cachedataset. Nothing went wrong when I use normal dataset.

[wyli]

thanks @xxsxxsxxs666 for reporting the issue, while I think the shape inconsistency could be addressed by using the EnsureSameShaped transform

MONAI/monai/apps/auto3dseg/transforms.py

Lines 25 to 32 in f9842c0

	class EnsureSameShaped(MapTransform):
	"""
	Checks if segmentation label images (in keys) have the same spatial shape as the main image (in source_key),
	and raise an error if the shapes are significantly different.
	If the shapes are only slightly different (within an allowed_shape_difference in each dim), then resize the label using
	nearest interpolation. This transform is designed to correct datasets with slight label shape mismatches.
	Generally image and segmentation label must have the same spatial shape, however some public datasets are having slight
	shape mismatches, which will cause potential crashes when calculating loss or metric functions.

I'm still not sure why this only happens with cacachedatset. Could you please share that test data if possible? I'm trying to reproduce the issue.

[xxsxxsxxs666]

Thanks for reply. It's very wierd that this problem only happen during training state, I can't recurrent that fault only using val dataset. But I still find something strange that shape of AI07_0012_01_B_diast_0.25mm.nii.gz after transform varies when the dataset's order is different. Sometime the shape is 639, but other times 640.Test code and data are in attachment. 19 熊小松 ***@***.***  

…

------------------ 原始邮件 ------------------ 发件人: "Project-MONAI/MONAI" ***@***.***>; 发送时间: 2023年2月6日(星期一) 下午5:39 ***@***.***>; 抄送: "19 ***@***.******@***.***>; 主题: Re: [Project-MONAI/MONAI] Why the data shape are different through transform "spacingd. (Issue #5935) thanks @xxsxxsxxs666 for reporting the issue, while I think the shape inconsistency could be addressed by using the EnsureSameShaped transform https://github.com/Project-MONAI/MONAI/blob/f9842c0dda70714c826eeb8b12bbf93d987a0584/monai/apps/auto3dseg/transforms.py#L25-L32 I'm still not sure why this only happens with cacachedatset. Could you please share that test data if possible? I'm trying to reproduce the issue. — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you were mentioned.Message ID: ***@***.***> 从QQ邮箱发来的超大附件 problem.7z (907.87M, 2023年03月09日 11:01 到期)进入下载页面：http://mail.qq.com/cgi-bin/ftnExs_download?t=exs_ftn_download&k=5a613130c24221a4ae1ff2764766514c070406530552075548535508074b575001041c095052524e045707035251075407020101616c6313170e535c040b4d541f610c&code=ea10afcc

[wyli]

thanks for the details @xxsxxsxxs666, I'm able to reproduce the issue and submitting a pull request #5950 to fix it.

The issue was with one of the images pixdim 0.346909 which was close to the target 0.35, the spacing transform correctly skipped the resampling but there's a buggy side-effect of assigning 0.346909 to self.pixdim. the assignment was not thread-safe and when used with the cachedataset with 4 workers, it showed some random behaviours, but your script narrowed it down, thanks!