Multi-site MRI harmonization with StarGAN
Making 3D T1 scans from four cohorts comparable enough to pool into one training set.
- Cohort
- 2,200 subjects
- Datasets
- 4
- Eval configs
- 8
Overview
An MRI volume carries the fingerprint of the scanner that produced it. Pool scans from several sites into one training set and a model can learn to identify the site instead of the pathology. This project trains a domain-conditional 3D StarGAN to map T1 volumes from NACC, OASIS3, and NIMHANS into an ADNI reference domain, across a 2,200-subject multi-site cohort.
The generator is a 3D residual network; the discriminator is a 5-layer PatchGAN with spectral normalization and gradient clipping. Adversarial training on full 3D volumes is unstable by default, and most of those architecture choices exist to make it converge at all.
Challenges
Image-fidelity metrics are conventionally reported on mid-slices, which is convenient and misleading — a generator can look excellent on the central slice and fail at the volume edges. PSNR, SSIM, mutual information, FID, and MMD are all computed over full volumes here instead.
Fidelity doesn't answer the question that matters. A harmonized volume that looks correct but erases the disease signal is worse than no harmonization, so downstream impact is measured directly: CN versus Dementia classification and brain-age regression across 8 harmonized and unharmonized train/test configurations.
The pipeline is being extended from a CN-only cohort to the full CN/MCI/Dementia cohort, where harmonization has to preserve signal across a much wider range of pathology.
Decisions
The calls that shaped this project, the alternatives they beat, and what each one cost.
Compute fidelity metrics over full volumes
over mid-slice metrics, as is conventional
WhyA 3D generator can look excellent on the central slice and fall apart at the volume edges. Mid-slice PSNR and SSIM hide exactly the failure mode that matters to a downstream 3D model.
Trade-offFull-volume PSNR, SSIM, mutual information, FID, and MMD are far more expensive to compute across every configuration.
Judge harmonization on downstream tasks
over image fidelity alone
WhyA harmonized volume that looks right but erases the disease signal is worse than no harmonization at all. CN versus Dementia classification and brain-age regression measure the thing that actually matters.
Trade-offEight harmonized and unharmonized train/test configurations to run instead of one fidelity number.
Spectral normalization and gradient clipping in the discriminator
over a standard PatchGAN discriminator
WhyAdversarial training on full 3D volumes is unstable by default — the discriminator overpowers the generator and training collapses. Constraining the discriminator's Lipschitz constant is what makes it converge at all.
Trade-offA deliberately weakened discriminator, which can slow how fast the generator improves.
Outcomes
In progress. The evaluation is designed so the result is interpretable either way — harmonization either improves downstream performance or it doesn't, and the 8 configurations show under which train/test splits.
TODO — results once the runs land.