cxx

VNet Tensorflow C++ Inference

C++ implementation of the V-Net architecture for medical image segmentation inferencing

Dependencies

Known good build dependencies:

• CMake 3.9.0

  • Following the CMake installation wizard to install CMake binary

• ITK 4.13.0

  • Please compile from source code with following CMake configurations

• Tensorflow 1.8.0 with C++ API

  • Please follow the CMake compilation instruction to build the C++ interface library. i.e. tensorflow_BUILD_SHARED_LIB options.

• Protobuf 3.5.0

  • The dependencies will be auto generated via Tensorflow C++ API superbuild.
  • Standalone build from source is possible but not recommended.

Build pass on Windows 10 with MSVC 2015. Test on your own on other platforms and compilers.

Build from source

1. Specify C++ source folder and target build directory with CMake (GUI/CCMake recommended)
2. Configure and provide necessary dependencies
3. Generate and Build

Prepare Tensorflow graph from python to C++

In python side we store the checkpoint in metagraph style for simple checkpoint loading. In C++ Tensorflow need to load graph and weight separately and we are now providing the meta_to_pb.py for the checkpoint conversion.

Points to note

• Currently the input and output only support absolute path in main.cxx
• We only illustrate one possible data type input with NIFTI format for convenience. It is possible to use JPG, TIFF or other image storage format.
• Only single batch inference is supported in present stage.
• Multi-threaded patch preparation is proposed for GPU utilization. This function is highly experimental and would like to request for development support.
• m_patchSize need to be same as the input placeholder size