Table 2.

Evaluation results for different choices of hyper-parameters, measured in terms of Segmentation Quality (SQ) and Recognition Quality (RQ), with training and inference times.

	SQ	RQ	Training (min.)	Inference (s)
Network depth
5	0.753	0.778	346	187
4	0.757	0.816	269	152
3	0.781	0.616	219	119
2	0.758	0.214 176	89
Loss
Weighted CE	0.757	0.816	269	152
Generalized dice	0.324	0.184	276	151
Focal	0.756	0.619	262	151
No border class	0.318	0.150	262	150
Tile size
256	0.769	0.647	79	250
384	0.767	0.756	169	216
512	0.757	0.816	269	152
524 without padding	0.766	0.729	227	4.5
768	0.769	0.733	332${}^{†}$	54${}^{†}$
Tile sampling
Area-based	0.757	0.816	269	152
Random	0.759	0.742	270	150
Fiber-centered	0.786	0.663	213	148
Proportional	0.783	0.675	213	150

Inference times are measured on image 18. CE denotes cross-entropy.

${}^{†}$Using a V100 GPU.