Automated segmentation of the dorsal root ganglia in MRI.
Journal:
NeuroImage
Published Date:
Apr 2, 2025
Abstract
The dorsal root ganglion (DRG) contains all primary sensory neurons, but its functional role in somatosensory and pain processing remains unclear. Recently, MR imaging techniques have been developed for objective in vivo observation of the DRG. In particular, DRG MR imaging endpoints such as DRG volume and DRG T2w signal are emerging as biomarkers with initial evidence of meaningful correlations with biochemical and genetic parameters as well as neuropathic pain as clinically relevant applications. However, the future validation and use of these novel imaging biomarkers critically depends on the development of fully automated methods for DRG image analysis. To date, DRG detection and evaluation on MR images has been limited to expert annotation through manual segmentation. Fast and operator-independent, yet accurate and robust, segmentation methods are required to enable observation of larger patient cohorts and across multiple sites. Thus, fully automated DRG segmentation is a prerequisite for the analysis of more complex microstructural and functional image datasets, such as from DRG diffusion tensor or perfusion metabolic imaging. Here, we developed a fully automated DRG segmentation workflow based on deep learning. A convolutional neural network (CNN) was trained using the nnU-Net framework on a large dataset of high-resolution 3D T2-weighted MR images of healthy controls (220 DRGs). Automated DRG segmentations generated with this network were on par with expert annotations (dice similarity coefficient of 0.87 for human expert vs. 0.89 for trained CNN) while being faster by a factor of 10. Finally, we validated the method in Fabry disease as a genetic model disorder for DRG pathomorphological injury. The trained CNN was able to reproduce the manually segmented changes known in FD patients as a function of FD genotype and FD pain phenotype. We developed a fully automated method for DRG MRI segmentation and validated its application as a novel imaging biomarker using the DRG injury example of Fabry disease.