Research map / Image generation / Image restoration
Towards Prospective Medical Image Reconstruction via Knowledge-Informed Dynamic Optimal Transport
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes KIDOT, a dynamic optimal transport framework whose transport cost and ODE dynamics embed imaging physics, trained on unpaired prospective measurements plus paired simulated data for MRI and CT reconstruction.
Problem
Supervised deep reconstruction trained on simulated (retrospective) measurement-image pairs degrades on real prospective data because the simulation misses real imaging knowledge; paired prospective data is scarce and often misaligned.
Approach
Model reconstruction as a continuous dynamic OT path from measurements to images, with an imaging-informed L1 data-fidelity transport cost integrated along the path and a gradient-flow ODE (A*(A(I)-I0) + learned regularizer gradient) as the transport equation, trained with a Kantorovich-Rubinstein critic plus supervised loss on paired simulated data.
Key results
- On simulated fastMRI knee 4x, KIDOT reaches 33.31 PSNR / 0.8518 SSIM / FID 10.68, above UAR (33.01 / 0.8425 / 19.56) and E2E-VarNet (32.56 / 0.8372 / 24.35).
- On real prospective T2-FLAIR 10x, KIDOT achieves FID 28.26 and KID×100 1.12, better than UAR (29.09 / 1.23) and PromptMR (29.52 / 1.46).
- On clinical prospective LDCT, KIDOT gets FID 22.25 vs. UAR 23.82 and ISTA-Net 24.86.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace UAR with KIDOT | Simulated MRI reconstruction · Simulated MRI (4x acceleration) | PSNR: 33.01 -> 33.31 | Helped |
| replace UAR with KIDOT | Real prospective MRI reconstruction · Real prospective MRI (10× acceleration) | FID: 29.09 -> 28.26 | Helped |
| replace UAR with KIDOT | Clinical prospective CT reconstruction · Clinical prospective CT (10-fold dose reduction) | FID: 23.82 -> 22.25 | Helped |
| add the potential network to supervised loss (ablation) | KIDOT on fastMRI knee · fastMRI knee (4x acceleration) | PSNR: 32.56 -> 32.92 | Helped |
| add the potential network to supervised loss (ablation) | KIDOT on fastMRI knee · fastMRI knee (4x acceleration) | SSIM: 0.8372 -> 0.8422 | Helped |
| add final-output physics consistency cost (ablation) | KIDOT on fastMRI knee · fastMRI knee (4x acceleration) | PSNR: 32.92 -> 33.01 | Helped |
| add final-output physics consistency cost (ablation) | KIDOT on fastMRI knee · fastMRI knee (4x acceleration) | SSIM: 0.8422 -> 0.8425 | Helped |
| apply physics consistency cost throughout the transport path instead of only at the final output (ablation) | KIDOT on fastMRI knee · fastMRI knee (4x acceleration) | PSNR: 33.01 -> 33.31 | Helped |
| apply physics consistency cost throughout the transport path instead of only at the final output (ablation) | KIDOT on fastMRI knee · fastMRI knee (4x acceleration) | SSIM: 0.8425 -> 0.8518 | Helped |
Surprising findings
- Static-OT and dynamic-OT baselines (OT-CycleGAN, FSBM) perform worse than the supervised E2E-VarNet on simulated MRI, despite learning from unpaired data.
Limitations
- Requires the known forward model A, which may be unavailable in practice.
- ODE-based transport with N steps raises computational cost relative to static methods.
- Real prospective data has spatial misalignment, so evaluation relies on distribution metrics (FID/KID) rather than pixel-wise metrics.
Experimental setup
Datasets: fastMRI multi-coil knee prospective United Imaging T2-FLAIR brain MRI prospective clinical abdominal LDCT/NDCT CT (UIH uCT960+) simulated low-dose CT from NDCT
Benchmarks: fastMRI knee simulated 4x real prospective T2-FLAIR MRI 10x clinical prospective LDCT CT
Metrics: PSNR SSIM FID KID
Baselines: CS-wavelet E2E-VarNet PromptMR OT-CycleGAN UAR FSBM DDS BM3D REDCNN ISTA-Net
Builds on: dynamic optimal transport Kantorovich-Rubinstein duality E2E-VarNet PromptMR ISTA-Net
Closest papers by meaning
- Knowledge Transfer Scaling Laws for 3D Medical Imaging (NeurIPS 2026 · similarity 0.86)
- ReMatching Dynamic Reconstruction Flow (ICLR 2025 · similarity 0.84)
- UGoDIT: Unsupervised Group Deep Image Prior Via Transferable Weights (NeurIPS 2025 · similarity 0.84)
- Residual-Conditioned Optimal Transport: Towards Structure-Preserving Unpaired and Paired Image Restoration (ICML 2024 · similarity 0.84)
- Expectile Regularization for Fast and Accurate Training of Neural Optimal Transport (NeurIPS 2024 · similarity 0.84)
- Spatiotemporal-Aware Neural Fields for Dynamic CT Reconstruction (AAAI 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.