P13. Semi-automatic, machine-learning based segmentation of peripheral nerves for quantitative morphometry: Comparison of low- and high-resolution MR neurography

Background Current state-of-the-art to diagnose peripheral neuropathy are neurological examination and electrodiagnostic studies. However, deeply situated nerves and plexus remain difficult to assess using these techniques, hence magnetic resonance neurography (MRN) emerged as a complementary method. MRN remains a qualitative approach and quantification, in terms of extraction of imaging biomarkers is needed to facilitate diagnosis and for follow-up examinations. Towards a multi-parametric quantitative imaging approach, accurate nerve segmentation needs to be performed first. Hampered by time-consuming manual annotation, we developed a semi-automatic, time-effective, hence clinically feasible segmentation approach based on machine-learning algorithms. In this study, we compared its performance on low- and high-resolution MRN. Materials and methods We acquired images of 9 healthy volunteers at the upper leg using 3 MRN sequences: lower resolution turbo inversion recovery magnitude (TIRM, voxel size of 78 × 78 × 4.0 mm 3 ), lower resolution spin echo T2-weighted (T2, voxel size .52 × .52 × 4.0 mm 3 ), and high resolution, fat-suppressed spin echo T2-weighted (hrT2, voxel size .25 × .25 × 3.3 mm 3 ). An expert manually segmented the sciatic nerve on the T2 and hrT2 images (ground truth, GT). Our segmentation algorithm is based on a decision forest with context- and intensity-based descriptors. We split the images into two sets: a multi-modal low-resolution (TIRM and T2), and a mono-modal high-resolution (hrT2). We perform a leave-one-out cross-validation for both sets and calculate the Dice coefficient to the manually annotated GT. Cross-sectional areas (CSA) were calculated from the middle slice of the image stack. Results Dice coefficients of 0.723 ± 0.202 and 0.735 ± 0.080 are achieved for the low- and high-resolution MRN images ( p  = 0.858, CI 95 %). Segmentation results and 3-D renderings of the same nerve segmented in low- and high-resolution are shown in Fig. 1 . CSA for GT were 34.0 ± 5.5 mm 2 (T2), and 35.4 ± 8.5 mm 2 (hrT2), using semi-automatic segmentation 24.8 ± 7.6 mm 2 (T2), and 25.7 ± 10.6 mm 2 (hrT2). Discussion and conclusion Semi-automatic, hence fast peripheral nerve segmentation is feasible in both low- and high-resolution MRN images of the upper leg, providing image-based quantitative morphometric data. The equal performance of the low-resolution images might be due to the multi-modal setting, which possibly provides more discriminative descriptors for nerve. The lower CSA using semi-automatic segmentation is due to more rigorous segmentation constraints. Future work will include the application of our method to localize and quantify nerve lesions in images of patients diagnosed with peripheral neuropathy.