Investigation of the influence of MR physics-based versus random intensity-based data synthesis methods for a generalizable spine segmentation network

Abstract

Recent advances in deep learning have greatly improved the automation of segmentation tasks. However, challenges remain in achieving robust performance on new domains, evidencing the need for large and diverse training datasets. In this study, two approaches, SynthMRI and SynthSeg, were implemented to generate new images during training, using available magnetic resonance (MR) scans, on a lumbar spine segmentation network. The main objective was to evaluate and compare their ability to generalize to unseen data. SynthMRI followed a physics-based approach that employed a set of quantitative MR images and Turbo Spin Echo (TSE) scans to synthesize new MR-like images with varied weightings using signal equations. In contrast, SynthSeg followed a domain randomisation strategy, where new images with random contrasts and intensities were generated from a set of anatomical label maps derived from TSE scans and the vertebrae segmentation by clustering image intensities. The evaluation of the predictions generated by the segmentation network trained with each approach revealed the ability of both SynthMRI and SynthSeg to generalize to images with unseen contrasts and patient populations. Specifically, SynthMRI achieved a mean Dice Similarity Coefficient (DSC) of 0.843 and a mean 95th percentile Hausdorff distance (HD95) of 3.712 mm, while SynthSeg obtained a mean DSC of 0.810 and a mean HD95 of 5.008 mm. Overall, no significant differences in performance were observed between the two methods. However, splitting the results by modality revealed that SynthMRI exhibited better performance than SynthSeg in TSE images. In conclusion, the outcomes of this study showed the great potential of both data synthesis strategies for achieving generalization in segmentation tasks.