550 images of level 1 were taken with a resolution of 300x1000. Then the path was manually labeled on the images.

These images can be used to train a variational auto encoder that learns the segmentation.