TABLE 1.

Ablation study results for clinical target volume (CTV) segmentation on three‐dimensional (3‐D) magnetic resonance (MR) images against expert manual segmentation

Method	Bladder localization	CTV span detection	CTV segmentation	DSC	p‐Value
1	–	2‐D Attention U‐Net	3‐D B2 U‐Net	*0.81 ± 0.08	0.0274
2	3‐D U‐Net	–	3‐D B2 U‐Net	*0.78 ± 0.05	1.81E − 08
3	3‐D U‐Net	2‐D Attention U‐Net	2‐D Attention U‐Net	*0.80 ± 0.05	7.17E − 05
4	3‐D U‐Net	2‐D Attention U‐Net	3‐D Dense U‐Net	*0.80 ± 0.10	0.0131
5	3‐D U‐Net	2‐D Attention U‐Net	3‐D U‐Net	*0.81 ± 0.08	0.0265
6	3‐D U‐Net	2‐D Attention U‐Net	3‐D B2 Dense U‐Net	*0.82 ± 0.05	0.0082
7	3‐D U‐Net	2‐D Attention U‐Net	3‐D B2 Attention U‐Net	*0.80 ± 0.06	1.25E − 04
Proposed	3‐D U‐Net	2‐D Attention U‐Net	3‐D B2 U‐Net	0.85 ± 0.03	–

Note: The asterisk denotes statistical significance using the t‐test performed to compare the average DSC value reported from our proposed method against those of alternatives.

Abbreviations: 2‐D, two‐dimensional; 3‐D, three‐dimensional; DSC, Dice similarity coefficient.