Generating large-scale highly heterogenous synthetic MRIs for robust spinal cord segmentation models

Motivation: Manually labelling large data sets is infeasible, especially heterogenous clinical scans to train segmentation MRI models. Generative models can mitigate this problem. Goal(s): Develop a generative model that introduces contrast and resolution variability into training datasets to segment heterogeneous clinical MRIs while reducing the need for manual labelling. Approach: We developed a generative model to synthesise spinal MRI images from label maps, applying intensity modelling, spatial deformations and noise to simulate anatomical and contrast variations for training. Results: We generated synthetic spinal MRIs and corresponding labels of varying contrasts and resolutions, showing higher similarity when mimicking original scans and lower when enabling contrast-randomisation. Impact: Our model paves the way for training contrast-agnostic and resolution-independent MRI segmentation models for spinal cord. This facilitates the processing of routine care data supporting more robust, translatable and generalisable models which can impact patients with neurological disorders.