. Author manuscript; available in PMC: 2020 Jul 10.

Published in final edited form as: IEEE Access. 2020 May 29;8:101550–101568. doi: 10.1109/access.2020.2998537

TABLE V.

Quantitative comparison of U-Net with Self-Training strategies in dentate and interposed nuclei segmentation.

Target	Dentate				Interposed

Metric	CMD (mm)	MSD (mm²)	DC	Volume (mm³)	CMD (mm)	MSD (mm²)	DC	Volume (mm³)
Metric	Manual labels	0.616±0.53	DC	0.431±0.13	0.853±0.06	724±184 (668±159)	DC	1.148±0.95	0.574±0.43	0.659±0.15	49±18 (53±24)
Self-training strategy #1 (pre-training)	1.45±1.25	0.770±0.18	0.698±0.12	1216±267 (668±159)	1.000±1.01	0.598±0.60	0.688±0.17	42±14 (53±24)
Self-training strategy #1 (fine-tune)	0.668±0.56	0.462±0.17	0.837±0.08	793±213 (668±159)	0.977±0.78	0.511±0.36	0.687±0.15	51±19 (53±24)
Self-training strategy #2 (model distillation)	0.684±0.61	0.452±0.13	0.843±0.06	806±190 (668±159)	1.02±0.93	0.52±0.38	0.693±0.17	48±20 (53±24)

Bold indicates p<0.05 for paired t-tests with U-Net (trained on manual labels). ( ) is ground truth volume.