PyTorchRadiomics

PyTorch implementation of PyRadiomics Extractor

Performance Improvement

It can speed up voxel-based features extraction significantly, especially GLCM features.

Using it to extract non-voxel-based features is NOT recommended (it is slower).

Voxel-based Features Extraction Performance Comparison

Intel i9-10900K v.s. RTX 3080 10G (dtype=torch.float64), Size=$16^3$

Type	CPU Time	Torch Time	Max Abs. Error	Max Rel. Error
GLCM	636s	23.8s	2.32e-09	7.92e-12
FirstOrder	4.3s	0.244s	2.84e-14	2.22e-16
GLRLM	1.71s	0.731s	2.72e-12	8.88e-16
NGTDM	4.03s	0.398s	3.27e-11	3.99e-15

Installation

pip install pytorchradiomics

Usage

Only two extra keyword arguments:

1. device: str or torch.device, default: "cuda"
2. dtype: torch.dtype, default: torch.float64

Direct usage:

from torchradiomics import (TorchRadiomicsFirstOrder, TorchRadiomicsGLCM,
                            TorchRadiomicsGLRLM, TorchRadiomicsNGTDM,
                            inject_torch_radiomics, restore_radiomics)

ext = TorchRadiomicsGLCM(
    img_norm, mask_norm,
    voxelBased=True, padDistance=kernel,
    kernelRadius=kernel, maskedKernel=False, voxelBatch=512,
    dtype=torch.float64, # it is default
    device="cuda:0",
    **get_default_settings())

features = ext.execute()

Or use injection to use RadiomicsFeatureExtractor:

from radiomics.featureextractor import RadiomicsFeatureExtractor
from torchradiomics import (TorchRadiomicsFirstOrder, TorchRadiomicsGLCM,
                            TorchRadiomicsGLRLM, TorchRadiomicsNGTDM,
                            inject_torch_radiomics, restore_radiomics)

inject_torch_radiomics() # replace cpu version with torch version

ext = RadiomicsFeatureExtractor(
    voxelBased=True, padDistance=kernel,
    kernelRadius=kernel, maskedKernel=False, voxelBatch=512,
    dtype=torch.float64, # it is default
    device="cuda:0",
    **get_default_settings())
ext.execute(img, mask, voxelBased=True)

restore_radiomics() # restore