pet-gmf

Generalized Mapping Function for PET image reconstruction

Table of Contents

• Introduction
• Installation
• Usage
• Contributing
• License

Introduction

This repository accompanies the paper "Learning Generalized Mapping Functions via Deep-Unrolling for PET Image Reconstruction" by Hitesh D. Khunti and Bhaskar D. Rao. We present a novel framework that unifies traditional iterative algorithms and modern deep learning techniques for PET image reconstruction. By leveraging a Generalized Mapping Function (GMF), we accelerate Expectation Maximization (EM) algorithms and reparametrize iterative methods as special cases of a learnable mapping between iterations. Our approach introduces a compact and interpretable deep neural network architecture that:

• Learns robust reconstructions with limited or no supervised data
• Performs voxel-wise acceleration of EM updates
• Demonstrates strong generalization to varying noise levels, scan durations, and out-of-distribution data.

Key results:

• Outperforms traditional EM and MAP algorithms in accuracy and speed
• Surpasses existing deep learning approaches in generalization and robustness
• Maintains theoretical interpretability with significant practical improvements Explore the codebase to reproduce experiments, train models in supervised and self-supervised settings, and test the GMF framework on synthetic and real-world PET data.

Installation

Prerequisites

• Install SIRF Superbuild docker

  • Install Docker and setup Nvidia drivers for docker access
  • Clone SIRF SuperBuild git repo
  • SIRF SuperBuild docker instructions
  • Clone this repository into a docker_shared folder.

     mkdir ~/docker_shared
     cd ~/docker_shared
     git clone [email protected]:SyneRBI/SIRF-SuperBuild.git
    

    This would allow code to be accessed both within the docker and your host

• Run SIRF docker Command template:

  cd ~/docker_shared/SIRF-SuperBuild/docker/
  sudo docker run \ 
  --name sirf-gpu1 \ 
  --rm -it -p 9999:8888 --gpus all \  
  -p 5022:22 \ 
  -v ./devel:/home/jovyan/work \ 
  -v ~/docker_shared:/home/jovyan/docker_shared \ 
  ghcr.io/synerbi/sirf:latest-gpu
  

• Setup docker and jovyan user settings

  • Login to the running docker

    sudo docker exec -it --user root sirf-gpu1 /bin/bash
    

  • Install vim editor, ssh server and run it so that you can ssh to the docker and connect via vs-code if required.

    	sudo apt-get update && \
    	sudo apt-get install openssh-server && \
    	sudo apt-get install vim && \
    	service ssh start
  

  • Set jovyan (default user) password passwd jovyan

• Install Pytroch and other required libraries

  pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121 && \
  pip3 install astra-toolbox==2.2.0 && \
  pip3 install pydicom && \ 
  pip3 install brainweb
  

• Connect to the docker now using ssh. Note the same forward port used as provided during docker run command in first step above.

  ssh jovyan@localhost -p 5022

• Clone pet-gmf repository

  cd ~/docker_shared
  git clone [email protected]:hdkhunti/pet-gmf.git
  

• Connect using vs-code to remote setup using ssh using the forward port 5022, jovyan username and set password.

• Configure vs-code to use /opt/conda/bin/python interpreter of the docker instance

Usage

• After the above installation steps, run pet-gmf/dnn_tir_main.py for trainning the DNN networks and pet-gmf/tir_algo_sweep_test.ipynb for running sweep tests on the trained networks and analytical algorithms.
• Use the following command to train gmf networks on brain phantom:

python dnn_tir_main.py --model gmf --nn_arch fcnn --epoch 100 --noise_pc 10.0  --phantom brain --lr=1e-5 --train_samps=1360 --test_samp=240 --val_samp=80 --en_mse_loss=1 --mdl_em_iter=4 --mdl_nlayers=8 --mdl_nchs=64 --mdl_act='selu' --mdl_t1byt2=0 

• Please cite the following paper if you use or build further on this work: H. D. Khunti and B. D. Rao, "Learning Generalized Mapping Functions via Deep-Unrolling for PET Image Reconstruction," in IEEE Transactions on Computational Imaging, vol. 11, pp. 1654-1667, 2025, doi: 10.1109/TCI.2025.3636751.

Contributing

Please reach out to hdkhunti in case you would like to contribute or collaborate further on this work.

License

This project is licensed under the MIT License — see the LICENSE file for details.