A deep learning-based self-adapting ensemble method for segmentation in gynecological brachytherapy

Radiation Oncology

Table 4 Summary of deep learning-based auto-segmentation results in gynecological cancer from other groups

Publication	Data type	Training cases	Testing cases	Method	Organ	DSC
Zhang et al. [48]	BT	73	18	DSD-UNET	Bladder	0.869 ± 0.032
					Rectum	0.821 ± 0.05
					HRCTV	0.829 ± 0.041
				3D-UNET	Bladder	0.802 ± 0.041
					Rectum	0.771 ± 0.062
					HRCTV	0.742 ± 0.062
Wang et al. [46]	EBRT	100	25	3D-CNN	Bladder	0.91 ± 0.06
					Rectum	0.81 ± 0.04
					HRCTV	0.86 ± 0.02
Liu et al. [49]	EBRT	77	14	Improved UNET	Bladder	0.924 ± 0.046
Liu et al. [49]	EBRT	77	14	Improved UNET	Rectum	0.791 ± 0.032
Rhee et al. [47]	BT	2254	140	CNN	Bladder	0.89 ± 0.09
					Rectum	0.81 ± 0.09
					HRCTV	0.86 ± 0.08
Our method	BT	205	30	nnU-NET	Bladder	0.936 ± 0.051
					Rectum	0.831 ± 0.074
					HRCTV	0.836 ± 0.07

If multiple network architectures are reported in the literature, the best-performing result was selected. The highest performance results (3D-Cascade) in our study were used for comparison. DSD-UNET: 3D-UNET incorporating residual connection, dilated convolution, and deep supervision

ISSN: 1748-717X