Skip to content

GPUfit Integration

DCEPrep uses GPUfit, a CUDA-accelerated curve fitting library, to dramatically speed up the numerical model fitting that underlies both T1 mapping and DCE pharmacokinetic analysis. GPUfit implements the Levenberg-Marquardt algorithm on the GPU and exposes MATLAB MEX bindings that are called directly by ROCKETSHIP.

What GPUfit Does in DCEPrep

GPU-based parallel fitting replaces CPU-bound iterative least-squares solvers. For large DCE datasets (hundreds of thousands of voxels), this can reduce fitting time from hours to minutes.

GPUfit is invoked at two points in the pipeline:

Pipeline step MEX file used Purpose
Preprocessing Step 10 (T1 mapping) GpufitCudaAvailableMex.mexa64 VFA-based T1 map generation via ROCKETSHIP
Analysis Step 1 (Ktrans fitting) GpufitConstrainedMex.mexa64 Extended Tofts model fitting via ROCKETSHIP's run_dce_cli

DCEPrep detects whether a GPU is available at runtime and records this in the per-case report. If GPUfit cannot find a CUDA-capable device it falls back to CPU fitting.

Pharmacokinetic Models Supported

GPUfit (petmri fork) includes GPU-optimised implementations of:

  • Extended Tofts (used by DCEPrep/ROCKETSHIP)
  • Tofts
  • Patlak
  • Tissue Uptake
  • Two-Compartment Exchange
  • T1 FA Exponential

How DCEPrep Locates GPUfit

DCEPrep searches for GPUfit's MEX files and MATLAB model definitions at startup:

# T1 mapping MEX (preprocess_all.sh)
GPUFIT_PATH=$(find $HOME -name 'GpufitCudaAvailableMex.mexa64' -printf '%h\n' -quit \
  || find / -name 'GpufitCudaAvailableMex.mexa64' -printf '%h\n' -quit)

# DCE analysis MEX (DCE_all.sh)
GPUFIT_PATH=$(find $HOME -name 'GpufitConstrainedMex.mexa64' -printf '%h\n' -quit \
  || find / -name 'GpufitConstrainedMex.mexa64' -printf '%h\n' -quit)

# MATLAB model ID definitions (both scripts)
GPUFIT_M_PATH=$(find $HOME -name 'ModelID.m' -printf '%h\n' -quit)

Both GPUFIT_PATH and GPUFIT_M_PATH are then passed to MATLAB via addpath:

addpath '$GPUFIT_PATH';
addpath '$GPUFIT_M_PATH';

In Docker, GPUfit is installed at /opt/Gpufit/ with MATLAB bindings at /opt/Gpufit/matlab64/, and the library path is set automatically:

LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/opt/Gpufit/matlab64/matlab

Installing GPUfit Without Docker

Download the prebuilt Linux release:

wget https://github.com/ironictoo/Gpufit/releases/download/1.3/Gpufit_1.3.0_linux.zip
unzip Gpufit_1.3.0_linux.zip -d /opt/Gpufit

Ensure the matlab64/ directory is reachable from $HOME so DCEPrep's path discovery can find the MEX files.

CUDA required

GPUfit requires a CUDA-capable NVIDIA GPU and a compatible CUDA toolkit. Without a GPU, ROCKETSHIP will fall back to CPU-based fitting, which is significantly slower on large datasets.

GPU Detection in Reports

DCEPrep's case report (case_report.py) checks the MATLAB fit logs for the string "Gpufit detected" and includes GPU availability in the per-subject QC report. This helps identify runs where GPU acceleration was unavailable and fitting may have taken longer or used different numerical paths.