MaGeZ contribution to the PETRIC reconstruction challenge

This repository contains algorithms developed by the MaGeZ team submitted to the 2024 PETRIC reconstruction challenge.

Authors

• Matthias Ehrhardt, Univeristy of Bath, United Kingdom
• Georg Schramm, KU Leuven, Belgium
• Zeljko Kereta, Univeristy College London, United Kingdom

Algorithms

ALG1 (main branch, ALG1 tag)

Better-preconditioned SVRG based on the works of:

• R. Twyman et al., "An Investigation of Stochastic Variance Reduction Algorithms for Relative Difference Penalized 3D PET Image Reconstruction," in IEEE TMI, 4, 2023, link
• J. Nuyts et al., ""A concave prior penalizing relative differences for maximum-a-posteriori reconstruction in emission tomography," in IEEE TNS, 49, 2022, link

The update can be summarized as:

$$ x^+ = x + \alpha P \tilde{\nabla} f(x) $$

where $\tilde{\nabla} f(x)$ is the SVRG stochastic gradient of a subset and $P$ is a diagonal preconditioner calculated as the harmonic mean of $x / A^T 1$ and the inverse of diagonal of the Hessian of the RDP, as proposed in work of Nuyts et al. The scalar step size $\alpha$ is set to 3 in the first epochs and then decreased to 1 and finally to 0.5 in late epochs.

ALG2 (ALG2 branch, ALG2 tag)

Similar to ALG1 but: The stepsize is set to 2.2 in the first epoch, and is then computed using the Barzilai-Borwein rule as described in Algorithm 2 in https://arxiv.org/abs/1605.04131, with m = 1 but using the short stepsize (see https://en.wikipedia.org/wiki/Barzilai-Borwein_method) adapted to preconditioned gradient ascent methods

ALG3 (ALG3 branch, ALG3 tag)

Using the same setup as in ALG2 but with two minor differences. First, we use a slightly smaller number of subsets. Second, we use a non-stochastic selection of subset indices by considering the cyclic group corresponding to the given number of subsets, finding all of its generators (i.e. the set of all coprimes of the number of subsets that are smaller), and then creating indices by consider specific generators at a time.

---

---

---

PETRIC: PET Rapid Image reconstruction Challenge

Participating

The organisers will provide GPU-enabled cloud runners which have access to larger private datasets for evaluation. To gain access, you must register. The organisers will then create a private team submission repository for you.

Layout

The organisers will import your submitted algorithm from main.py and then run & evaluate it. Please create this file! See the example main_*.py files for inspiration.

SIRF, CIL, and CUDA are already installed (using synerbi/sirf). Additional dependencies may be specified via apt.txt, environment.yml, and/or requirements.txt.

• (required) main.py: must define a class Submission(cil.optimisation.algorithms.Algorithm) and a (potentially empty) list of submission_callbacks, e.g.:

  • main_BSREM.py
  • main_ISTA.py
  • main_OSEM.py

• apt.txt: passed to apt install

• environment.yml: passed to conda install, e.g.:

  name: winning-submission
  channels: [conda-forge, nvidia]
  dependencies:
  - cupy
  - cuda-version =11.8
  - pip
  - pip:
    - git+https://github.com/MyResearchGroup/prize-winning-algos

• requirements.txt: passed to pip install, e.g.:

  cupy-cuda11x
  git+https://github.com/MyResearchGroup/prize-winning-algos

Tip

You probably should create either an environment.yml or requirements.txt file (but not both).

You can also find some example notebooks here which should help you with your development:

• https://github.com/SyneRBI/SIRF-Contribs/blob/master/src/notebooks/BSREM_illustration.ipynb

Organiser setup

The organisers will execute (after installing nvidia-docker & downloading https://petric.tomography.stfc.ac.uk/data/ to /path/to/data):

# 1. git clone & cd to your submission repository
# 2. mount `.` to container `/workdir`:
docker run --rm -it --gpus all -p 6006:6006 \
  -v /path/to/data:/mnt/share/petric:ro \
  -v .:/workdir -w /workdir synerbi/sirf:edge-gpu /bin/bash
# 3. install metrics & GPU libraries
conda install monai tensorboard tensorboardx jupytext cudatoolkit=11.8
pip uninstall torch # monai installs pytorch (CPU), so remove it
pip install tensorflow[and-cuda]==2.14  # last to support cu118
pip install torch --index-url https://download.pytorch.org/whl/cu118
pip install git+https://github.com/TomographicImaging/Hackathon-000-Stochastic-QualityMetrics
# 4. optionally, conda/pip/apt install environment.yml/requirements.txt/apt.txt
# 5. run your submission
python petric.py &
# 6. optionally, serve logs at <http://localhost:6006>
tensorboard --bind_all --port 6006 --logdir ./output

FAQ

See the wiki/Home and wiki/FAQ for more info.

Tip

petric.py will effectively execute:

from main import Submission, submission_callbacks  # your submission (`main.py`)
from petric import data, metrics  # our data & evaluation
assert issubclass(Submission, cil.optimisation.algorithms.Algorithm)
Submission(data).run(numpy.inf, callbacks=metrics + submission_callbacks)

Warning

To avoid timing out (currently 10 min runtime, will likely be increased a bit for the final evaluation after submissions close), please disable any debugging/plotting code before submitting! This includes removing any progress/logging from submission_callbacks and any debugging from Submission.__init__.

• data to test/train your Algorithms is available at https://petric.tomography.stfc.ac.uk/data/ and is likely to grow (more info to follow soon)
  • fewer datasets will be available during the submission phase, but more will be available for the final evaluation after submissions close
  • please contact us if you'd like to contribute your own public datasets!
• metrics are calculated by class QualityMetrics within petric.py
  • this does not contribute to your runtime limit
  • effectively, only Submission(data).run(np.inf, callbacks=submission_callbacks) is timed
• when using the temporary leaderboard, it is best to:
  • change Horizontal Axis to Relative
  • untick Ignore outliers in chart scaling
  • see the wiki for details

Any modifications to petric.py are ignored.