How to handle subvolumes?

[mattwarkentin]

Hi @fepegar,

I am not sure if torchio has support for this, or if I need to do some pre-processing outside of torchio first. Basically, I have CT and binary mask file pairs. The masks were of course derived from the CT scan, but when the mask is saved to disk, for memory reasons, only the minimal bounding box containing the ROI is saved. The affine/origin is updated accordingly, so the sub-volume is "aware" it is contained within the CT, but from torchio's perspective these are two volumes of different spatial shape and with different affines (at least the origin component).

Is there any good way to re-align these two volumes?

For additional context:

• CT: 512x512x322, spacing=[0.74, 0.74, 1], origin=[186.1, 333.1, -363.0]
• Mask: 35x35x26, spacing=[0.74, 0.74, 1], origin=[-49.8, 86.8, -111.0]

Hoping there is a good way to geometrically align based on their origins, and then pad the mask with 0's until it matches the spatial shape of the CT, and then updating them to have the same origin/affine.

[fepegar]

Hey @mattwarkentin.

1. Are you compressing those masks? The file size should be just some KB
2. You can use Resample('ct') to match the FOVs

[fepegar]

No compression.

Any good reason for this? Just curious.

[mattwarkentin]

Good reasons for not compressing? Or good reasons for storing the masks as minimal bounding boxes?

[fepegar]

I guess it's the kind of the same question. My point is that if you compressed the masks, the file size would be quite small and you wouldn't need to care about the bounding boxes and use Resample. On the other hand, using smaller and uncompressed FOVs would optimize data loading, which I guess is better if you have a lot of storage but want things to be very fast.

[mattwarkentin]

It's actually just a consequence of the particular software used for doing the ROI segmentations. By default, it saves masks as minimal bounding boxes. I would've preferred it maintained the full geometry and just used compression instead (or nothing at all, the file sizes aren't that large compared to the CT).

[fepegar]

Thanks!

[romainVala]

Hi there, I think compression if of course very efficient (especially for binary mask with a lot of zeros), so you spare space on the disk but not on memory when you load those volumes.

I think this issue is related to #734 , it is exactly the same (sub volume) but the reason is that otherwise it does not fit into ram memory