To develop a method for nerve morphometric analysis through deep learning.

Most neurological diseases produce one of two key pathological changes – axonal loss or demyelination – or a combination of the two. Therefore, studying these disorders requires rigorous quantification of myelin and axon pathology. Traditional manual quantification is time-consuming and may suffer from inter-observer variation. Deep-learning has been utilized to automate image analysis. The aim of the present study is to develop a Convolutional Neural Network (CNN) – based approach to segment images of mouse nerve.

We used Keras, a deep-learning library, to create a CNN based on U-net architecture for improved localization of image features. Training data included 280 microscopic images of mouse sciatic nerve cross-sections paired with their respective segmentation masks obtained in previous studies of neuropathy mouse models.

After training, accuracy plateaued at 0.91 dice coefficient and the validation dice coefficient varied between 0.81 and 0.85. Compared to the manual method, the CNN-based automated method exhibited a 2.5% decrease of nerve fiber density, 4.2% lower axonal diameter, 2.0% larger myelin thickness, and 2.6% lower G-ratio. Distribution of myelinated fiber diameters was very similar between the two methods, thus no size of nerve fibers was disproportionately affected by the automation. After training, measurements took 16-20 minutes per image while manual segmentation took 65-76 minutes.