How hopkins-cig/gtp works
GTP is a research codebase for Bayesian medical-image reconstruction using cross-modality translation priors, alongside conventional diffusion and analytic/iterative baselines. Its user workflow runs MR-to-CT or CT-to-PET experiments from prepared data and checkpoints, applies a reconstruction method, and saves results for quantitative comparison. The README documents the experiment families and asset layout; sampled source confirms the diffusion model, CT/PET physics operators, and configuration utilities. Experiment-script internals were not sampled, so links from experiment runners to specific algorithms are limited to documented capabilities.
Main entry points: mr_ct MR-to-CT runs · ct_pet CT-to-PET runs · train.py Model training · ddbm.py Translation model
Components 17
Reconstruction experiments
- MR-to-CT runs
scripts/run_testing/mr_ct - CT-to-PET runs
scripts/run_testing/ct_pet - Reconstruction baselines
scripts/run_testing - Model training
scripts/train.py
Generative priors
- Translation model
src/models/ddbm.py - Bridge diffusion
src/models/karras_diffusion.py - Denoising network
src/models/edm_unet.py - Image encoder
src/models/encode_sam.py
Imaging physics
- CT projection
src/physics/pbct.py - PET projection
src/physics/pet.py
Inference support
- Inference configuration
src/inference_config.py - Experiment settings
configs/testing - Config loader
configs/testing/__init__.py
Other components
- Researcher
- Prepared images
- Model checkpoints
- Reconstruction results
Connections 26
- Researcher → MR-to-CT runs: runs
- Researcher → CT-to-PET runs: runs
- MR-to-CT runs → Inference configuration: resolves assets
- CT-to-PET runs → Inference configuration: resolves assets
- MR-to-CT runs → Config loader: loads settings
- CT-to-PET runs → Config loader: loads settings
- Config loader → Experiment settings: reads
- MR-to-CT runs → Prepared images: loads
- CT-to-PET runs → Prepared images: loads
- MR-to-CT runs → Model checkpoints: loads
- CT-to-PET runs → Model checkpoints: loads
- MR-to-CT runs → Translation model: uses translation prior
- CT-to-PET runs → Translation model: uses translation prior
- Translation model → Bridge diffusion: denoises with
- Bridge diffusion → Denoising network: calls
- MR-to-CT runs → CT projection: projects CT measurements
- CT-to-PET runs → PET projection: projects PET measurements
- MR-to-CT runs → Reconstruction baselines: supports baseline runs
- CT-to-PET runs → Reconstruction baselines: supports OSEM and DDS
- MR-to-CT runs → Reconstruction results: writes
- CT-to-PET runs → Reconstruction results: writes
- Researcher → Model training: trains
- Model training → Translation model: trains model
- Model training → Prepared images: reads
- Model training → Model checkpoints: writes
- MR-to-CT runs → Image encoder: uses unpaired encoding