Unsupervised Medical-Image Registration
An unsupervised VoxelMorph-based deep model for 3D liver MRI image registration, trained without ground-truth alignments.
Python PyTorch VoxelMorph Medical Imaging
Overview
An unsupervised deep learning model for 3D liver MRI image registration, built on the VoxelMorph architecture during MSc research at Otto-Von-Guericke University. The model learns a deformation field directly from unlabelled scan pairs, removing the need for manually annotated ground-truth alignments.
Key Features
- Preprocessing pipeline that normalizes and resamples scan volumes to centre liver and tumour regions
- Unsupervised VoxelMorph-based registration network trained without ground-truth deformation fields
- Evaluation against Jaccard and Dice overlap metrics on held-out liver volumes
Results
Achieved a Jaccard coefficient of 91.44 and a Dice score of 93.27 on 3D liver MRI registration.