Towards improving edge quality using combinatorial optimization and a novel skeletonize algorithm

BMC Medical Imaging

Table 3 Evaluation results using optimal parameters on the test part of each data set

Data set	Work	Post-processing	SDE	\({\varvec{\tilde{x}}_{\varvec{SDE}}}\)	IoU-box	\({\tilde{\varvec{x}}_{{{\mathbf{IOU-{box}}}}}}\)
Kidney boundaries	[14]\(^{\text{b}}\)	Original\(^{\text{d}}\)	232.9	103.5	0.32	0.33
Kidney boundaries	[14]\(^{\text{b}}\)	Our method	2515.0	954.5	0.04	0.02
NYU depth dataset V2	[29]\(^{\text{a}}\)	Original	4.9	4.3	0.6	0.6
NYU depth dataset V2	[29]\(^{\text{a}}\)	Our method	4.2	3.8	0.6	0.6
	[31]\(^{\text{a}}\)	Original	8.1	7.3	0.9	1
	[31]\(^{\text{a}}\)	Our method	7.4	6.4	1	1
BSDS 500	[25]\(^{\text{a}}\)	Original	50.0	25.6	0.8	0.8
	[25]\(^{\text{a}}\)	Our method	21.8	12.0	0.6	0.6
	[32]\(^{\text{c}}\)	Original	11.1	7.9	0.5	0.6
	[32]\(^{\text{c}}\)	Our method	12.0	6.8	0.8	0.8
	[33]\(^{\text{c}}\)	Original	9.5	7.1	0.5	0.4
	[33]\(^{\text{c}}\)	Our method	10.3	6.3	0.8	0.9

Reported metrics are mean and median values (\({\tilde{x}}\)). We compare the results of Algorithm 1 to each method specific results. We note that the original post-processing implementations did not include a thresholding step except for [14]
Originally used post-processing methods: \(^{\text{a}}\) 2D skeletonize algorithm, \(^{\text{b}}\) 3D skeletonize algorithm, \(^{\text{c}}\) standard non-maximum suppression. \(^{\text{d}}\) Values in this row were taken from the original paper and not recomputed (c.f., discussion section point three for details)

ISSN: 1471-2342