PET Reconstruction with L-BFGS-B

A Python wrapper that adapts SIRF objective functions for use with scipy's L-BFGS-B optimiser. Some of the code here should be generalisable but currently it's relatively hard-coded to deal with RDP regularised PET reconstruction

Core Implementation

The main functionality is in two files:

• pet_recon/optimisation/objective_function.py - Wraps SIRF objectives, handles the sign flip (return -objective_func(image) - see below)
• pet_recon/optimisation/bfgs_optimiser.py - Implements the L-BFGS-B optimisation loop with proper bounds and callbacks

Note:

• SIRF objective functions are maximisers (log-likelihood), but L-BFGS-B is a minimiser.
• CUDA RelativeDifferencePrior is giving NaNs for some reason.

Quick Start

python scripts/main.py --data_path /path/to/pet_data --num_iters 100 --gpu

There's a run.sh script that I'm using to run stuff locally

Requirements

• SIRF (PET reconstruction framework)
• scipy (L-BFGS-B optimiser)
• numpy

Why L-BFGS-B?

• Handles non-negativity constraints naturally via bounds
• Memory efficient for large PET images
• Well-tested, robust convergence
• Built-in line search

The rest of the codebase is just organisational scaffolding around these core optimisation components.

Some results

Recontruction with 50 iterations of L-BFGS-B for NEMA phantom with RDP