SSMI 2026 hackathon projects

[casperdcl]

Suggestions for the SSMI 2026 hackathon, 5 Oct 2026.

For inspiration, see also issues/PRs/discussions in https://github.com/UCL/STIR

Tip

See #1683 for examples of how to write a suggestion in this discussion.

See also the hackathon channel on the synerbi Discord server.

[casperdcl]

MRI: providing SIRF+XNAT as alternative reconstruction for clinical data

Leads: @paskino, Emily, @ckolbPTB

• test of reconstruction pipeline on phantom data, which shows hallucinations on the Siemens AI-recon tool
• establish fixes to the raw data for import into SIRF. What we have to do and why.
• generate DICOM files with appropriate metadata (currently empty and/or not tested)
• prepare code to run reconstruction in XNAT
• test code in a XNAT pipeline
• Use code carbon to assess energy consumption

[casperdcl]

Can we convert CUDA code to AMD

Leads: If anyone available?

• Would need small CUDA project to get started. STIR does have a limited amount of CUDA, so might be feasible.
• Possible access to AMD GPUs (STFC has AMD W6600)

[gschramm]

Would be very interested in this. unfortunately I cannot join on 05 Oct (teaching + remote presentation).

However, I can offer to spend half a day on testing whether I can get libparallelproj running on AMD GPUs - if someone can get me access to those for 1 day.

[bwynneHEP]

I can provide some example code and container recipes

A colleague made this git repo: https://github.com/StewMH/gpu_translate

[bwynneHEP]

Monte Carlo simulations

Leads: @bwynneHEP

• Reading of Monte Carlo generated data into STIR
  • GATE 10 support (PR Add generic multi-level repeater support for GATE10 / LAFOV scanners (Quadra) #1765)
  • Make GATE IO friendlier
  • Explore SimSET tools for STIR (WIP PRs Simset input #1678 and Update simset conversion files to read out multiple energy bin sinograms #1608)
  • Update https://github.com/UCL/STIR-GATE-Connection?

[ruaridhg]

Measuring energy consumption on different computing platforms

Leads: Charlotte, @ruaridhg

• How to measure energy consumption (metrics):
  • CodeCarbon (giving energy consumption for different bits of the computing chain. uses nvidia-tools, but also monitors CPU). Can run via CLI.
  • GreenAlgorithms (takes input from user on hardware) (additional possibility)
  • Investigate how this works in VMs, dockers etc
  • Choose which metrics for comparisons
• STIR:
  • Install on both x86 and ARM-based architecture
  • Dockerised version?
  • Candidate algorithms: OSEM and PETRIC2 winner
• Resources (Ruaridh and Charlotte will use UCL resources):
  • Cricket NVIDIA A100 GPU
  • Locust NVIDIA GH200 Grace Hopper Superchip
  • DGX Spark NVIDIA GB10 Grace Blackwell w/ STIR installed on ARM-based architecture
  • Myriad A100s/V100s w/ STIR installed x86 architecture
  • UCL-AWS (EC2 instance with NVIDIA H100 and x86 architecture)
  • Cloud access will be provided by STFC

[bwynneHEP]

Possibly relevant to discussion of cloud computing: https://arxiv.org/pdf/2505.08530

[vcosmi]

Making STIR/SIRF more efficient

Leads: @vcosmi, @KrisThielemans

• Profiling Tools

  • Explore profiling tools to identify execution bottlenecks (NVIDIA and other alternatives: https://hackingcpp.com/cpp/tools/profilers, https://dleggat.github.io/gppt-ws01-talk/presentation_v0#/title-slide)

• CUDA-Managed Pointers Integration

  • Review and test the CUDA managed memory approach used to modify the memory allocation for selected STIR classes (PR CUDA Managed Pointers integration for selected components #1768)

• SIRF Compatability

  • Test whether the memory allocation changes break or seamlessly interface with SIRF

• SPECT GPU SPECTGPU SPECTGPU #1687

• @casperdcl enable self-hosted CUDA runner in CI #1764

• Resources

  • University/Cloud access for development

• Data to Download

  • some SPECT phantom data can be found here https://osf.io/t69jg/overview, it can be downloaded via terminal as:
    osf -p t69jg clone
unzip t69jg/osfstorage/NPL-SPECT-demo.zip -d SPECT/Organs

[CatrionaWim]

Dynamic PET reconstruction: balancing energy consumption, accuracy and precision

Lead: @CatrionaWim

We will use SIRF to reconstruct dynamic PET data and use the energy consumption tools from the Measuring energy consumption topic measure energy use. We will compare reconstructed images, time–activity curves and kinetic estimates with simulated ground truth.

We will provide some prepared data and a code package for different reconstruction configurations in the SIRF container.