Segmentation of Peripheral Nerves From Magnetic Resonance Neurography: A Fully-Automatic, Deep Learning-Based Approach

Abstract

Diagnosis of peripheral neuropathies relies on neurological examinations, electrodiagnostic studies, and since recently magnetic resonance neurography (MRN). The aim of this study was to develop and evaluate a fully-automatic segmentation method of peripheral nerves of the thigh. T2-weighted sequences without fat suppression acquired on a 3 T MR scanner were retrospectively analyzed in 10 healthy volunteers and 42 patients suffering from clinically and electrophysiologically diagnosed sciatic neuropathy. A fully-convolutional neural network was developed to segment the MRN images into peripheral nerve and background tissues. The performance of the method was compared to manual inter-rater segmentation variability. The proposed method yielded Dice coefficients of 0.859 ± 0.061 and 0.719 ± 0.128, Hausdorff distances of 13.9 ± 26.6 and 12.4 ± 12.1 mm, and volumetric similarities of 0.930 ± 0.054 and 0.897 ± 0.109, for the healthy volunteer and patient cohorts, respectively. The complete segmentation process requires less than one second, which is a significant decrease to manual segmentation with an average duration of 19 ± 8 min. Considering cross-sectional area or signal intensity of the segmented nerves, focal and extended lesions might be detected. Such analyses could be used as biomarker for lesion burden, or serve as volume of interest for further quantitative MRN techniques. We demonstrated that fully-automatic segmentation of healthy and neuropathic sciatic nerves can be performed from standard MRN images with good accuracy and in a clinically feasible time.

Keywords: health; machine learning; magnetic resonance imaging; magnetic resonance neurography; peripheral nervous system diseases; sciatic nerve; segmentation.

Figure 1

Overview of the proposed peripheral nerve segmentation method. (A) Training of the neural network with T2 and ground truth image slices, and (B) testing of the trained neural network yields a segmentation of the peripheral nerve without the need of a ground truth.

Figure 2

Segmentation evaluation metrics of our method (Auto-GT) compared to inter-rater variability (R-R). Boxplot of the (A) Dice coefficient, (B) Hausdorff distance, (C) volume similarity, and (D) segmentation time separated by healthy volunteer and patient cohort. *The segmentation time of our method is less than 1 s and therefore barely visible in the boxplot.

Figure 3

Segmentation of the sciatic nerve of a patient. (Left) 3-dimensional rendering of the segmentation. The color map encodes the surface-to-surface distance of the segmentation to the ground truth. (Right) The segmentation boundaries (green) and ground truth boundaries (blue) on the T2 image are shown for three slices along the nerve course.

Figure 4

Interpretability of the Hausdorff distance (HD) metric for peripheral nerves. (A) The same nerves segmented by the three raters (left to right) are depicted in red. One rater does not segment all branches (arrows), which results in a large HD. (B) The consensus ground truth (left) compared to the segmentation results by our method (right). A falsely segmented vein (arrow) by our method results in a large HD.

Figure 5

Potential of computer-assisted segmentation of peripheral nerves for imaging biomarkers. (Left column) 3-dimensional renderings of the sciatic nerve with lesions colored red: (Top row) a healthy volunteer, (Middle row) a patient with a focal lesion, (Bottom row) and a patient with an extended lesion. (Middle column) Cross-sectional area evolution and (Right column) lesion burden evolution obtained from the segmentation could be used as biomarkers to assess disease severity and progression, or to categorize the lesion type. Note that not all peripheral nerve lesion types show morphometric abnormalities, hence a combination with signal intensity (or other quantifiable MR parameters) is necessary to assess the lesion burden. The quantified signal intensity evolution was assessed by segmenting hyperintense nerve fascicle bundles on a co-registered T2-weighted sequence with fat suppression using inversion recovery.