Table 2.

Segmentation and biometry performance of different deep learning models

Methods	DSC (%)	DF (mm)	AD (mm)	HD (mm)
U-Net	97.72 ± 1.42	1.83 ± 2.58	2.36 ± 2.11	1.35 ± 0.85
V-Net	97.85 ± 1.32	0.95 ± 2.58	2.01 ± 1.87	1.30 ± 0.75
Attention V-Net	97.91 ± 1.24	− 0.57 ± 2.46	1.85 ± 1.71	1.28 ± 0.79
DS U-Net	97.85 ± 1.26	− 1.06 ± 2.48	2.00 ± 1.80	1.31 ± 0.78
DS V-Net	97.87 ± 1.27	− 0.38 ± 2.53	1.89 ± 1.73	1.29 ± 0.74
DAG V-Net	97.93 ± 1.25	0.09 ± 2.45	1.77 ± 1.70	1.27 ± 0.80

Data are presented as mean ± standard deviation. The segmentation and biometry performance data were obtained after uploading the experimental results by different methods to the official assessment system of the HC18 Challenge. The DAG V-Net proposed in this work achieved the best performance among the six deep learning methods (data indicated in bold)

DSC Dice similarity coefficient, HD Hausdorff distance, DF head circumference difference, AD head circumference absolute difference, DS deeply supervised, DAG deeply supervised attention-gated