# Best Practices for Custom Image Filters in ITK with Multi-Threading Support
**URL:** https://discourse.itk.org/t/best-practices-for-custom-image-filters-in-itk-with-multi-threading-support/7547
**Category:** Uncategorized
**Created:** [May 28, 2025, 11:16am UTC](https://discourse.itk.org/t/best-practices-for-custom-image-filters-in-itk-with-multi-threading-support/7547 "2025-05-28T11:16:46Z")
**Posts on this page:** 2
**Page:** 1
### Author:  [@Kelvin](https://discourse.itk.org/u/Kelvin)
#### Post date: [May 28, 2025, 11:16am UTC](https://discourse.itk.org/t/best-practices-for-custom-image-filters-in-itk-with-multi-threading-support/7547/1 "2025-05-28T11:16:46Z")
Hello everyone,
I’m currently working on developing a custom image filter in ITK for a research project, and I’m hoping to get some guidance on best practices, especially when it comes to performance and thread safety.
My filter performs a voxel-wise computation that’s not easily expressed through existing filters like UnaryGenerator or BinaryFunctor. The operation involves accessing a local neighborhood for each pixel and applying a custom mathematical transformation that depends on both intensity values and spatial relationships. Because of the spatial dependencies, I’m unsure how to structure the filter to best take advantage of ITK’s multi-threading.
I’ve reviewed the ITK Software Guide’s sections on creating custom filters, and I understand the general pattern of overriding ThreadedGenerateData.
My questions are:
- Are there recommended examples or templates for building multi-threaded, neighborhood-aware filters in modern ITK (v5.x or newer)?
- Is there a preferred strategy for handling temporary or intermediate buffers that need to be shared across threads but written in a thread-safe way?
> [@Was itk::simple::AdaptiveHistogramEqualizationImageFilter ever multi-threaded?](https://discourse.itk.org/t/was-itk-adaptivehistogramequalizationimagefilter-ever-multi-threaded/3424):
>
> Would like to run this filter on some real big images: ~ 8K x 10K x 2K voxels. Have a cluster and can multi-thread as much as necessary. Per: [https://itk.org/pipermail/community/2015-October/010010.html](https://itk.org/pipermail/community/2015-October/010010.html), was this filter ever multi-threaded? Alternatively - if I wanted to run the filter on overlapping blocks of the input in separate processes, would specifying the block overlap radius equal to the filter radius guarantee the same output as if I had run on the whole image? I.e. are there any ed…
- Would it be better to split the operation into two filters — one for gathering statistics and one for applying the transformation?
- Any tips for leveraging the ThreadedImageRegionPartitioner or other recent multi-threading improvements?
However, while exploring [sap abap training in chennai](https://www.igmguru.com/erp-training/sap-abap-training-in-chennai), I’m also concerned about memory access patterns—particularly when using neighborhood iterators—and how to avoid race conditions when writing intermediate results or shared statistics.
I’d appreciate any examples, insights, or gotchas the community can share. Thanks in advance!
---
### Author:  [@matt.mccormick](https://discourse.itk.org/u/matt.mccormick)
#### Post date: [May 28, 2025, 3:33pm UTC](https://discourse.itk.org/t/best-practices-for-custom-image-filters-in-itk-with-multi-threading-support/7547/2 "2025-05-28T15:33:09Z")
Hello @Kelvin ,
For building multi-threaded code with modern ITK,
- [`itk::MultiThreaderBase::ParallelizeImageRegion`](https://docs.itk.org/projects/doxygen/en/v5.4.0/classitk_1_1MultiThreaderBase.html#a911f06d8531fb7f9ef33715eb9e87e8a) is quite nice. The `ThreadingFunctorType` is a `std::function`
For working with image neighborhoods, I highly recommend @N-Dekker’s [`itk::ShapedImageNeighbohoodRange`](https://docs.itk.org/projects/doxygen/en/v5.4.0/classitk_1_1ShapedImageNeighborhoodRange.html). This is fast with an excellent modern, C++ interface.
- [example](https://examples.itk.org/src/core/common/imagebufferandindexrange/documentation)
- [example with lambda](https://github.com/InsightSoftwareConsortium/ITK/blob/380e8116d5e64b4b064aa9a18fe052af487243c5/Modules/Core/Common/test/itkShapedImageNeighborhoodRangeGTest.cxx#L684-L686)
Regarding storage, we do require modern C++ standards for that include functionality like `thread_local`.