Table 3 Comparison of Dice scores for urinary system segmentation between the proposed modified U-Net model, U-net with various components, and 2D based models in state of the art with simple multi-class implementation.

	Avg	Kidney	Ureter				Urinary bladder
	Avg	Kidney	Total ureter	Proximal ureter	Middle ureter	Distal ureter	Urinary bladder
Proposed modified U-net	0.8623 (± 0.0289)	0.9613 (± 0.0044)	0.7225 (± 0.1584)	0.7121 (± 0.1286)	0.6558 (± 0.1228)	0.6085 (± 0.1235)	0.9032 (± 0.0579)
U-Net model
Baseline	0.7835	0.8751	0.57	0.5875	0.5427	0.5515	0.9055
With transposed convolution	0.8021	0.907	0.6012	0.6124	0.595	0.6004	0.8982
With multi-segmentation decoder	0.815	0.9315	0.5998	0.6134	0.5885	0.598	0.9137
With ROI extraction and slice windowing	0.831	0.9376	0.6309	0.6385	0.6163	0.6005	0.9243
Other 2D based models in state of the art
UNet++	0.7964	0.914	0.5774	0.5888	0.5698	0.5589	0.898
PraNet	0.8208	0.9392	0.6501	0.6762	0.6487	0.6087	0.8732
nnU-Net	0.8332	0.9423	0.6573	0.6838	0.631	0.6163	0.8999
Swin-UNet	0.8481	0.9488	0.6924	0.6974	0.6462	0.673	0.9031

Quick links

Search