Deep-learning based segmentation of challenging myelin sheaths

The segmentation of axons and myelin in electron microscopy images allows neurologists to highlight the density of axons and the thickness of the myelin surrounding them. These properties are of great interest for preventing and anticipating white matter diseases. This task is generally performed manually, which is a long and tedious process.We present an update of the methods used to compute that segmentation via machine learning. Our model is based on the architecture of the U-Net network. Our main contribution consists in using transfer learning in the encoder part of the UNet network, as well as test time augmentation when segmenting. We use the SE-Resnet50 backbone weights which was pre-trained on the ImageNet 2012 dataset.We used a data set of 23 images with the corresponding segmented masks, which also was challenging due to its extremely small size. The results show very encouraging performances compared to the state-of-the-art with an average precision of 92% on the test images. It is also important to note that the available samples were taken from elderly mices in the corpus callosum. This represented an additional difficulty, compared to related works that had samples taken from the spinal cord or the optic nerve of healthy individuals, with better contours and less debris.