Back to projects

Learn2Breath (Learn2Reg 2026 Challenge)

A training-free lung CT registration pipeline, accepted at Learn2Reg 2026 (MICCAI) with an invited oral presentation.

SWE (AI/MLOps)Medical ImagingImage RegistrationComputer VisionPyTorchDocker

Ships as a single, fully deterministic Docker image with all model weights baked in at build time, verified to run end to end with --network=none. A fixed three-argument entrypoint writes a channel-first NIfTI displacement field, and the pipeline degrades gracefully through several fallback tiers so it always exits cleanly with a valid output, even under segmentation failure or tight VRAM. End-to-end runtime measured at 69.3s per pair on a Tesla T4, comfortably inside the evaluation budget.

Learn2Reg 2026: Medical Image Registration Challenge at MICCAI 2026, Strasbourg
Organizer invitation to present: an oral presentation and poster slot at Learn2Reg 2026

What I built

  • Packaged the full pipeline into a single Docker image with all TotalSegmentator weights baked in at build time, verified to run end to end with --network=none for the evaluation environment's no-internet constraint.
  • Defined a fixed three-argument ENTRYPOINT contract (fixed inspiration CT, moving expiration CT, output path) writing a channel-first float32 NIfTI displacement field on the fixed image's grid at 1.5mm spacing.
  • Layered the pipeline with fallbacks so it always exits cleanly with a valid, correctly shaped output: MIND-SSC-only registration with an HU-threshold lung box if segmentation fails or returns fewer than 3 of 5 lobes, one retry if that also raises, and a zero displacement field as a last resort.
  • Added a pre-flight VRAM check that shrinks the discrete search radius automatically when less than 10GB is free, so the container can't fail with an out-of-memory error on constrained hardware.
  • Kept the full pipeline deterministic end to end, with the Adam instance optimisation initialised from the deterministic coupled-convex solution and run for a fixed iteration count, and bounded the folding-repair stage with a 30s wall-clock cap.
  • Measured 69.3 +/- 1.3s end-to-end runtime per pair on a Tesla T4 (53.1s segmentation, 7.5s registration, 1.4s folding repair), with a 72.3s worst case across the 10 validation pairs.

Tech stack

SWE (AI/MLOps)Medical ImagingImage RegistrationComputer VisionPyTorchDocker

Key results

0.973

Lobar DSC (5 lobes)

0.0035%

%NDV (post-repair)

69.3s

Runtime per pair (T4)

Accepted, invited oral

Status

Back to all projects