Skip to content

architecture.md

Latest commit

 

History

History
63 lines (48 loc) · 2.11 KB

architecture.md

File metadata and controls

63 lines (48 loc) · 2.11 KB

ITK Architecture

Module Structure

ITK uses a module-based architecture. Each module is a self-contained unit with declared dependencies via itk-module.cmake:

itk_module(ITKCommon
  DEPENDS ITKEigen3 ITKVNL
  PRIVATE_DEPENDS ITKDoubleConversion
  COMPILE_DEPENDS ITKKWSys
  TEST_DEPENDS ITKTestKernel ITKMesh
  DESCRIPTION "Core ITK classes..."
)

Modules are organized in:

  • Modules/Core/ — Image, Region, Point, Vector, pipeline infrastructure
  • Modules/Filtering/ — Image processing filters
  • Modules/IO/ — JPEG, PNG, NIFTI, DICOM, and other format I/O
  • Modules/Registration/ — Image registration algorithms
  • Modules/Segmentation/ — Segmentation algorithms
  • Modules/Numerics/ — Optimization and numerical methods
  • Modules/ThirdParty/ — Vendored deps (HDF5, JPEG, TIFF, VNL, Eigen3); managed as git subtrees
  • Modules/Bridge/ — Bridges to VTK, NumPy
  • Modules/Remote/ — External modules fetched at build time

Template-Heavy C++ Design

ITK uses C++ templates for dimension-agnostic (2D/3D/nD) and pixel-type-agnostic code. Both are compile-time parameters:

itk::Image<TPixel, VImageDimension>

Compile-time polymorphism is achieved via traits and policy classes rather than virtual dispatch.

Pipeline Architecture

ITK uses a lazy evaluation pipeline:

  • Process objects (itk::ProcessObject) — filters that transform data
  • Data objects (itk::DataObject) — itk::Image, meshes, etc.
  • Filters connect via SetInput() / GetOutput() and execute only when Update() is called
  • itk::SmartPointer<T> handles all memory management — never call delete on ITK objects

Python Wrapping

Wrappers are generated via SWIG. Not all C++ types are wrapped. To check available instantiations, look in wrapping/ subdirectories of each module.

Common wrapped pixel types: F (float), D (double), UC (unsigned char), US (unsigned short).

Snake_case Python API mirrors C++ (e.g., itk.median_image_filter()).

import itk
image = itk.imread('input.png')
smoothed = itk.median_image_filter(image, radius=2)
itk.imwrite(smoothed, 'output.png')