Transform-CropToMask: crops image to the extremes of a bounding box about a given mask

[ashkanpakzad]

🚀 Feature

Transform that crops given image to the extremes of a bounding box about a given mask.

Motivation

Such a transform is really useful in lung imaging for example, where we cut out the rest of the chest CT (which can include the unwanted abdomen) to crop to a lung segmentation that is fairly easy to acquire.

Pitch

A transform CropToMask that takes mask_name and a padding variable and crops an input ScalarImage to the bounding box about mask_name with padding to the crop bounding box.

If the padding value exceeds the limits of the ScalarImage then pad the output as provided by padding_mode.

Alternatives

As suggested by @fepegar could also make this an implementation of CropOrPad that takes target_shape=None and mask_name.

Additional context

Currently, I implement this by preprocessing my data separately. Example for lung mask in chest CT.

The CT images provided from QIN LUNG CT (https://wiki.cancerimagingarchive.net/display/Public/QIN+LUNG+CT#b8d88cce4fd14620bef4e5e35ec3d589) under Creative Commons Attribution 3.0 Unported License (https://creativecommons.org/licenses/by/3.0/). The citations are:

• Goldgof, Dmitry, Hall, Lawrence, Hawkins, Samuel, Schabath, Matthew, Stringfield, Olya, Garcia, Alberto, … Gillies, Robert. (2015). Data From QIN_LUNG_CT. The Cancer Imaging Archive. http://doi.org/10.7937/K9/TCIA.2015.NPGZYZBZ
• Jayashree Kalpathy-Cramer, Sandy Napel, Dmitry Goldgof, Binsheng Zhao. QIN multi-site collection of Lung CT data with Nodule Segmentations. http://dx.doi.org/10.7937/K9/TCIA.2015.1BUVFJR7
• Clark K, Vendt B, Smith K, Freymann J, Kirby J, Koppel P, Moore S, Phillips S, Maffitt D, Pringle M, Tarbox L, Prior F. The Cancer Imaging Archive (TCIA): Maintaining and Operating a Public Information Repository, Journal of Digital Imaging, Volume 26, Number 6, December, 2013, pp 1045-1057. (https://doi.org/10.1007/s10278-013-9622-7)

[fepegar]

Thanks, @ashkanpakzad. If different volumes in the dataset end up having different shapes, it won't be possible to use a batch size > 1. Are you planning to train with patches?

[ashkanpakzad]

Perhaps I've not thought this through for my current project where I'm using whole 3D images. CropOrPad should already do the job withtarget_shape necessary to achieve batch size>1.

Though in a general preprocessing strategy, it would be useful for patches as you suggest.

[fepegar]

Perhaps I've not thought this through for my current project where I'm using whole 3D images. CropOrPad should already do the job withtarget_shape necessary to achieve batch size>1.

My suspicions are confirmed, then :)

Assuming your spacings are all the same, you could use as target_shape the largest size across all masks in your dataset and then use mask_name so the FOV is always centered on each mask.

[ashkanpakzad]

Yes, very good point, thank you. I'll park it here for the future maybe :)

[fepegar]

Thanks anyways, for being happy to contribute!

[fepegar]

Reopening after more interest has been shown on #677.