Swin Transformers learn to align 3D brain scans without labels

What it does TransMorph is a PyTorch implementation of a 2022 Medical Image Analysis paper that applies Vision and Swin Transformers to unsupervised 3D medical image registration. It learns to warp and align volumetric scans—primarily brain MRIs—without relying on manually annotated ground truth. The repo includes four model flavors: a standard hybrid Transformer-ConvNet, two diffeomorphic variants that enforce topology-preserving deformations, and a Bayesian version that estimates registration uncertainty.

The interesting bit Instead of treating transformers as a drop-in replacement, the project bundles a full experimental suite: it ships with training and inference scripts for eight classical and learning-based baselines (from ANTsPy to VoxelMorph) and four competing Transformer architectures so you can compare long-range attention against old-school CNNs and optimization-based methods on the same footing. The Bayesian variant is a nice touch—it doesn’t just warp the image; it tells you how much to trust the warp.

Key highlights

• Four model variants: standard, diffeomorphic (probabilistic and B-spline), and Bayesian uncertainty estimation.
• Supports both mono-modal and multi-modal registration with six similarity losses (NCC, MI, SSIM, MIND-SSC, etc.) and four deformation regularizers.
• Includes pre-trained models and reproducible training pipelines for public IXI and OASIS brain MRI datasets.
• Bundles baseline implementations for direct comparison with classical methods (SyN, NiftyReg, LDDMM) and modern CNN/Transformer competitors.
• A Docker image is available for out-of-the-box brain MRI registration.

Caveats

• The README’s claims of holding first place on the MICCAI 2021 L2R challenge date from 2022; current standings are unclear.
• The authors’ original JHU brain MRI and Duke CT data are not distributed due to restrictions, so you will need to source your own scans or rely on the provided public datasets.
• The repo is a research artifact tied to a specific paper, and the authors now point to a newer lightweight package (MIR) for broader deployment.

Verdict Worth a look if you are building or benchmarking deep-learning registration pipelines and need a well-documented Transformer baseline with uncertainty and diffeomorphic options. Skip it if you want a generic, framework-agnostic tool rather than a paper-specific PyTorch implementation.