Potential overflow with very large niftis

[ofgulban]

Hi @neurolabusc ,

I have noticed that e.g. -Tmean flag results in no output when input is a very large nifti (>10GB). I have also noticed that I can do operations on two very large niftis (e.g. niimath file1.nii.gz -sub file2.nii.gz output.nii.gz) but the resulting nifti voxels in the output nifti are not correct beyond 2^16 th voxel. Therefore, I am suspecting an overflow error.

My niimath version v1.0.20220720 Clang12.0.5 (64-bit MacOS)

[neurolabusc]

Thanks for the report. Note that v1.0.20220720 is ancient, and more recent versions exit without creating any output beyond an error code. However, the most recent commit to the wasm branch (commit aeab04d, v1.0.20260713) supports huge images and will be in the next release.

Supported on huge images: voxelwise math (-add/-sub/-mul/-div, unary math), thresholding (-thr/-thrp/-clamp/…), temporal reductions (-Tmean/-Tstd/-Tmax/…), and Gaussian smoothing (-s). So the -Tmean case from your report now works. Some operations that I do not believe are appropriate for huge images return a "not huge-image capable" message. In future, these could be supported if there is a need.

Caveats: Processing a huge image needs enough RAM to hold it (≥ ~9 GB for an 8.6 GB volume, more for multi-step chains). The WebAssembly build is unchanged here — wasm32 has a 4 GB address ceiling, so it keeps the INT_MAX limit by design.

I would strongly recommend upgrading to a recent version of niimath. You will see performance optimizations and new features including:

• Fast linear image registration that be used for defacing
• Diffuson tensor properties
• Ability to generate bitmaps

[ofgulban]

Thanks @neurolabusc ,
I have tried -Tmean now with niimath version v1.0.20260703 Clang17.0.0 BSD (64-bit MacOS). I had the same no output issue. Looking forward to try the next release.