Table 1:

Quantitative comparison of different segmentation models. All models are variants of that in Fig. 2. ‘Seg’ denotes a single decoder segmentation network trained using a supervised loss; ‘Seg + PH’ denotes fine-tuning a pretrained segmentation network using the topological loss, but directly on the segmentation outputs; ‘Seg + Cyl’ denotes the network in Fig. 2, which is trained using only supervised losses; ‘Seg + Cyl + PH’ denotes the proposed method. For every metric, the mean and standard deviation are presented. Refer to the text for the explanation on the evaluation metrics. P-values are computed by conducting paired t-tests between the baseline method and the other methods with the Dice coefficients.

Method	Dice	HD (mm)	HD95 (mm)	ASD (mm)	p-value
Seg [4]	0.838 ± 0.044	58.894 ± 20.423	16.081 ± 6.886	2.733 ± 0.879	-
Seg + PH [5]	0.822 ± 0.061	64.754 ± 28.825	14.666 ± 6.612	2.802 ± 1.076	0.205
Seg + Cyl	0.839 ± 0.048	63.968 ± 22.219	14.192 ± 5.494	2.644 ± 0.840	0.679
Seg + Cyl + PH	0.852 ± 0.045	51.642 ± 9.292	12.180 ± 6.232	2.330 ± 0.764	0.032