[SyneRBI/SIRF-Contribs] Hackathon projects (Discussion #31)
Casper da Costa-Luis
Suggestions for the AIRBI hackathon that will take place on 11 & 12 March 2026.
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you are subscribed to this thread.
Kris Thielemans
Project 1: Merge pytorch support framework to SIRF
Leads: @casperdcl @KrisThielemans
Idea: See SyneRBI/SIRF#1305. This PR is essentially ready, but throws up some errors in CI. We hope to complete this before the hackathon, but...
We could (possibly trivially) also update it to change as_array() calls to asarray() to avoid data copies.
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you are subscribed to this thread.
Christoph Kolbitsch
Project 2: Proof-of-concept for MR DL-reconstruction
Leads: @ckolbPTB
Idea: Include the SIRF MR acquisition model into a deep inverse DL reconstruction.
Starting points:
- MR brainweb SIRF acquisition model: https://github.com/SyneRBI/SIRF-Exercises/blob/master/notebooks/Introductory/acquisition_model_mr_pet_ct.ipynb
- handling of 3D brainweb data in deep inverse: https://deepinv.github.io/deepinv/auto_examples/physics/demo_mri_tour.html#explore-3d-mri
- MODL-based image reconstruction: https://deepinv.github.io/deepinv/auto_examples/self-supervised-learning/demo_artifact2artifact.html#define-model
- Linear physics operator: https://deepinv.github.io/deepinv/api/stubs/deepinv.physics.LinearPhysics.html
Main limitation: No GPU-support for MR acquisition model available yet in SIRF
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you are subscribed to this thread.
Matthias J. Ehrhardt
Project 3: Plug and play regularization for PET
Leads: @mehrhardt
Idea: Include trained regularizers from DeepInverse into PET reconstruction in SIRF.
Starting points:
- Use pretrained convex ridge regularizer or input convex neural network in either gradient-based or primal-dual PET reconstruction.
- Use pretrained denoiser to replace proximal operator in either gradient-based or primal-dual PET reconstruction.
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you are subscribed to this thread.
Matthias J. Ehrhardt
Project 4: Deep Image Prior for PET
Leads: @mehrhardt TBC
Idea: Use deep image prior for PET reconstruction.
- This will need to have a PET forward model in DeepInverse and could be a good use-case.
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you are subscribed to this thread.
Christoph Kolbitsch
Project 5: Evaluation of MR DL-reconstruction
Idea: Evaluate artefacts on commercial DL-reconstruction by comparing it to SIRF/CIL reconstructions.
Background: DL reconstruction provided by vendor leads to artefacts in the reconstructed knee images which are mistaken for pathologies. Patients had to be recalled for a second imaging session. SIRF/CIL reconstruction could help to understand where these artefacts come from and maybe even avoid recalling patients.
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you are subscribed to this thread.
Andrew Wang
Project 6: DeepInverse wrapper for SIRF/CIL physics operators
Leads: @Andrewwango, @tachella
Idea: DeepInverse abstracts imaging forward operators into a physics, allowing them to be used with all other modern DL tooling in DeepInverse e.g. pretrained models, diffusion sampling, self-supervised learning. We should create a physics that wraps SIRF/CIL operators (e.g. PET acquisition), so that users can use this physics natively with DeepInverse algorithms and tools.
Starting point:
The wrapper could follow our Astra integration TomographyWithAstra and look something like:
class deepinv.physics.EmissionTomographyWithSIRF(deepinv.physics.LinearPhysics): ...
Tasks:
- Allowing efficient device transfer when objects are on CUDA
- Write end-to-end example, like Low-dose CT with ASTRA backend and Total-Variation (TV) prior
- Unit test physics, ensuring adjointness and normalisation
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you are subscribed to this thread.
Andrew Wang
Project 7: DeepInverse modality-specific utils (generators, IO, vis...)
Leads: @Andrewwango, @tachella
Idea: DeepInverse abstracts imaging forward operators into a physics, allowing them to be used with all other modern DL tooling in DeepInverse e.g. pretrained models, diffusion sampling, self-supervised learning. We should create a physics that wraps SIRF/CIL operators (e.g. PET acquisition), so that users can use this physics natively with DeepInverse algorithms and tools.
Starting point:
The wrapper could follow our Astra integration TomographyWithAstra and look something like:
class deepinv.physics.EmissionTomographyWithSIRF(deepinv.physics.LinearPhysics): ...
Tasks:
- Allowing efficient device transfer when objects are on CUDA
- Write end-to-end example, like Low-dose CT with ASTRA backend and Total-Variation (TV) prior
- Unit test physics, ensuring adjointness and normalisation
New to DeepInverse? Tutorial: 5-min quickstart tutorial, Tutorial: how to contribute
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you are subscribed to this thread.
Kris Thielemans
Project 8: update STIR data to CUDA managed pointers
Leads: @casperdcl @KrisThielemans @evgueni-ovtchinnikov
Idea: Follow-up from UCL/STIR#1683 (comment).
needed for Project 6 #31 (comment)
- likely need to complete the STIR work
- adapt
sirf.STIRdata-containers (images first) - expose underlying CUDA managed pointer to python
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you are subscribed to this thread.
MaxTousss
Projet 9: Enable the use of DeepInverse as plug-in in CASToR
Leads: @sstute1
Idea: CASToR enables, among other things, reconstruction of PET list-mode data. By construction, it is ill adapted to be interfaced inside Python. However, we could use python module as a plug-in for image-space processing, thus making it possible to use deep-learning models in key steps of a reconstruction scheme.
Tasks:
- Validate the capability of using python plug-in in CASToR, including pytorch with GPU
- Implements an algorithm with a pre-learned deep-learning component (e.g., denoiser)
- Implements an algorithm with deep-learning component that is trained in the reconstruction process (e.g., DIP)
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you are subscribed to this thread.
Biguri
Project 10: Self supervised frameworks for real data in tomography (XCT, Neutron, etc)
Leads: @AnderBiguri @MargaretDuff
Idea: The LION toolbox (work in progress) implements pre-made classes and scripts for reproducing papers in learned tomography reconstruction. We are currently working on exploring self-supervised methods with high success. In particular we have implementations for Noise2Inverse, Noisier2Inverse, Equivariant2Inverse, Proj2Proj, Sparse2Inverse (and SURE_pg, but doesn't really work well).
Example script for Noise2Inverse here. Other examples to come soon, but essentially its just changing the solver class.
These methods have been rarely used with real data (only recently a bit in the Equivariant2Inverse paper with the 2DeteCT datast). Using them for real tomographic data (of any time) would be interesting. LION supports operators of any type, but only CT is implemented now.
Tasks:
- Finish cleaning up/implementing the Noise2Inverse variants
- Find real data datasets to test the models on (e.g. https://zenodo.org/records/17250237, but others welcomed)
- Evaluate performance for different scenarios
- Evaluate generalization capabilities and/or transfer learning (important for inclusion in real-world environments)
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you are subscribed to this thread.
Biguri
Of course, we could also chat about how to connect LION-DeepInverse-SIRF more too! Unfortunately I am unlikely to be able to engage with the hackathon a lot due to tons of other pressing tasks, but I think this is something that could benefit all of us!
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you are subscribed to this thread.
Andrew Wang
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you are subscribed to this thread.