Single-axon level automatic segmentation and feature extraction from immuhistochemical images of peripheral nerves

Abstract

<jats:title>Abstract</jats:title><jats:p>Quantitative descriptions of the morphology and structure of peripheral nerves is central in the development of bioelectronic devices interfacing the nerves. While histological procedures and microscopy techniques yield high-resolution detailed images of individual axons, automated methods to extract relevant information at the single-axon level are not widely available. We implemented a segmentation algorithm that allows for subsequent feature extraction in immunohistochemistry (IHC) images of peripheral nerves at the single fiber scale. These features include short and long cross-sectional diameters, area, perimeter, thickness of surrounding myelin and polar coordinates of single axons within a nerve or nerve fascicle. We evaluated the performance of our algorithm using manually annotated IHC images of 27 fascicles of the swine cervical vagus; the accuracy of single-axon detection was 82%, and of the classification of fiber myelination was 89%.</jats:p>

Organ- and function-specific anatomical organization of the vagus nerve supports fascicular vagus nerve stimulation

“…However, when it comes to nerve anatomical structures and segmentation, the applications of deep-learning methods are limited to EM [57] or ultrasound images [61]. For IHC imaging, segmentation efforts on nerve fibers have relied on manual or semi-automatic processes [62, 63], while newer approaches have used neural networks to segment and reconstruct fascicular organization in peripheral nerves [64]. All these approaches are limited to fascicle segmentation and characterization, not taking advantage of all the staining and imaging capabilities of IHC imaging.…”

“…The use of deep-learning based algorithms on anatomically guided, medical image segmentation has received Despite the increased interest in deep-learning based algorithms for anatomically guided, medical image segmentation [62], when it comes to nerve anatomical structures and segmentation, the applications of deep-learning are limited to EM [57] and ultrasound images [61]. In IHC imaging, segmentation of nerve fibers has relied on manual or semi-automated processes [62,63]; more recently, neural networks have been used to segment and reconstruct fascicles in peripheral nerves [64]. In micro-CT and other optical tomography imaging modalities, software-aided manual segmentation is the standard in processing of fascicles and nerve structures [1,65,66].…”

“…After testing different computer vision frameworks [42], we optimized the quantitative anatomy pipeline by segmenting nerve structures on micro-CT and IHC images using a mask-RCNN deep-learning architecture [43]. Deep-learning based algorithms have recently received attention in anatomically guided, medical image segmentation [44].…”

“…To the best of our knowledge, this is the first methodology that allows complete characterization of the main functional fiber types in the vagus, including small, unmyelinated fibers using light microscopy. The use of deep-learning based algorithms on anatomically guided, medical image segmentation has received Despite the increased interest in deep-learning based algorithms for anatomically guided, medical image segmentation [62], when it comes to nerve anatomical structures and segmentation, the applications of deep-learning are limited to EM [57] and ultrasound images [61]. In IHC imaging, segmentation of nerve fibers has relied on manual or semi-automated processes [62, 63]; more recently, neural networks have been used to segment and reconstruct fascicles in peripheral nerves [64].…”

“…We designed a set of algorithms and a data pipeline to segment and extract nerve, fascicle and single-axon level features from IHC images of the vagus. First, we developed a fiber segmentation and feature extraction algorithm by combining standard computer vision methods with statistical modeling and geometrical manipulations, including Gaussian mixture models and Voronoi tessellation, and achieved a single-axon detection accuracy of 82% and myelination classification accuracy of 89% [62] To improve the detection accuracy, we further trained a Mask R-CNN deep convolutional neural network on axon instance segmentation. The model was pretrained on the COCO Instance Segmentation task [72].…”