torch-mando: Differentiable Domain Transform in PyTorch based on MandoCT

This library introduces Differentiable Forward Projection (FPJ) and Filtered Back Projection (FBP) in PyTorch to enable dual-domain deep learning and optimization.

Now support:

• equidistant fan-beam
• geometric calibration

Installation

• Check the prerequisites

  • torch-mando only works on Linux.
  • Prepare CUDA compiler nvcc, PyTorch, and GCC compiler.

• Clone the repository

  git clone [email protected]:CandleHouse/torch-mando.git --depth 1
  cd torch-mando

• Install by running

  python setup.py install
  

About Geometry

Now support the following "parameter.jsonc" files to reduce the influence of geometry on the final results:

• PMatrixFile
• SIDFile
• SDDFile
• ScanAngleFile
• DetectorOffcenterFile

⚠️ It is advised to directly use pMatrix file instead of other files, since pMatrix along is able to describe all possible geometric deformations. You can even use torch-mando to reconstruct fan beam projections or project fan beam images with geometric vibrations, and the result is almost the same with Mandoct.

cfg = MandoFanBeamConfig(...)
# add params file
cfg.addPmatrixFile(readParamsFile('./params/pmatrix_file.jsonc'), pmatrixDetEltSize=0.4)
cfg.addSIDFile(readParamsFile('./params/sid_file.jsonc'))
cfg.addSDDFile(readParamsFile('./params/sdd_file.jsonc'))
cfg.addScanAngleFile(readParamsFile('./params/scan_angle.jsonc'))
cfg.addDetectorOffCenterFile(readParamsFile('./params/offcenter_file.jsonc'))

Example

The main idea is to assemble a Config object which includes all necessary settings to perform FPJ and FBP, and pass it to MandoFanbeamFpj / MandoFanbeamFbp along with the image / sinogram.

cfg = MandoFanBeamConfig(imgDim=imgDim, pixelSize=imgPixelSize, sid=750, sdd=1250, detEltCount=detEleCount,
                         detEltSize=detEleSize, views=views, reconKernelEnum=KERNEL_RAMP, reconKernelParam=1, fovCrop=False)
# In your network
# If you prefer nn.functional F style
def forward(self, x):
   x = MandoFanbeamFbp(x, cfg)

# If you prefer nn.Module layer style
def __init__(self) -> None:
   self.fbpLayer = MandoFanbeamFbpLayer(cfg)
def forward(self, x):
   x = self.fbpLayer(x)

View example for code details.

Declaration

This work highly depends on previous works by ustcfdm/mangoct, njjixu/mangoct, SEU-CT-Recon/mandoct and matteo-ronchetti/torch-radon. This work draws lots of lessons from them. Thanks for their wonderful work.

Cite this

Please use the following BibTex to cite this work, or click Cite this repository on the right.

@software{torch_mando,
  author = {SEU-CT-Recon},
  license = {MIT},
  title = {torch-mando: Differentiable Domain Transform in PyTorch based on MandoCT},
  url = {https://github.com/SEU-CT-Recon/torch-mando}
}

License

MIT