I work on a very similar task in a large research program (on part is to segment carbon fibers in carbon fiber reinforced concrete). However, my task is a little easier than yours, as I have less fibers to segment than you.

Your initiall approach is correct: manual labelling. I know, this is a pain. I did 4 volumes using 512x512x20 voxels and then applied heavy offline augmentation for volumes using this project . In my case, it worked very well (visually). You can pm me, if you want to know more and see some results.

I used dragonfly by ORS to create the masked volume. They provide a non commercial license and the segmentation tool works pretty well (see their youtube playlist). You can also use their ai feature but I cannot tell you how well that works, cause I used the previous mentioned project.