CT → MRI Synthesis (3D CycleGAN)

OverviewCopy link to section

During my internship with the MRI group at BioMaps, I studied unpaired, fully volumetric CT-to-UTE-MRI translation of the thorax. The practical goal was label transfer: generate MRI-like volumes from annotated CT scans while keeping the original anatomy and voxel geometry intact, so that scarce MRI segmentation data can be supplemented without spatially warping the labels.

What I builtCopy link to section

• A reproducible 3D preprocessing pipeline covering orientation unification, 1 mm isotropic resampling, shape-safe padding, and modality-specific intensity normalization.
• A controlled comparison of 3D CycleGAN variants using ResNet-9b, U-Net++, and an adapted DC-CycleGAN backbone.
• Experiments with LSGAN and WGAN-GP objectives, cycle consistency, and structure-aware losses.
• Hist-CycleGAN, which adds a differentiable histogram loss to align the generated and real MRI intensity distributions during training.

Main findingCopy link to section

The models generally preserved anatomy and volumetric coherence, but realistic UTE-MRI contrast was the harder problem. Standard objectives produced global intensity shifts, scale drift, saturation, or mode collapse even when cycle-reconstruction scores were strong. This made the intensity-distribution gap—not geometric fidelity—the main bottleneck.

The histogram-aware objective improved both distributional metrics while retaining anatomical fidelity. Residual contrast bias remains, so the work points toward region-conditioned histogram supervision, CDF-based distances, MRI noise modelling, and downstream validation on vascular-tree segmentation.

ReportCopy link to section

The full report documents the datasets, preprocessing safeguards, architectures, ablations, qualitative comparisons, and training details.

Read the full CT-to-MRI synthesis report (PDF)