Computer-assisted counting of retinal cells by automatic segmentation after TV denoising

BMC Ophthalmology

Table 2 Overall cell counts by different methods

No.	Image	Manual count #	Direct segm.	TV denoising & segm.	m /c /f
			# / rel. error	# / rel. error
	Orangutan
1	OrangWhM	2328	1784 / 24.9	2286 / 1.8	30 / 2.0 / 5
2	OrUinf-0001pr	1192	793 / 33.5	1168 / 2.0	30 / 3.0 / 5
	Domestic cat
3	loStack	557	518 / 7.0	543 / 2.5	30 / 2.5 / 5
4	luStack	552	584 / 5.8	562 / 1.8	30 / 2.5 / 5
5	roStack	536	539 / 0.6	520 / 3.0	30 / 2.5 / 5
6	ruStack	584	573 / 1.9	573 / 1.9	30 / 2.5 / 5
	Manul
7	bcergcomp	128	110 / 14.1	128 / 0.0	20 / 2.5 / 25
	Lynx
8	Movie-17-1	708	675 / 4.7	659 / 6.9	30 / 2.5 / 5
	Cheetah
9	1+5CSCStack	85	84 / 1.2	91 / 7.1	30 / 3.5 / 40
10	Stack SCFoc.	687	656 / 4.5	634 / 7.7	30 / 2.5 / 5
	Jaguar
11	Combination	289 (S-cones)	177 / 38.8	275 / 4.8	30 / 3.5 / 5
	Pangolin
12	492_20_09	369	109 / 70.5	367 / 0.5	20 / 2.5 / 25
	Agouti
13	A4scal	43	14 / 67.4	44 / 2.3	40 / 2.5 / 150
14	A5scal	48	20 / 58.3	54 / 12.5	40 / 2.5 / 150
15	A6scal	52	44 / 15.4	53 / 1.9	40 / 2.5 / 150
16	B1scal	31	35 / 12.9	43 / 38.7	40 / 2.5 / 150
17	B2scal	55	51 / 7.3	56 / 1.8	40 / 2.5 / 150
18	C4scal	213 (?)	95 / 55.4	234 / 9.8	30 / 2.5 / 50
	Mean error		23.6	5.9
	Standard dev.		23.6	8.6

Within Algorithm 1, the following parameters have been chosen: The median is taken over a field of m×m pixels, the TV denoising procedure is applied with regularization parameter α = 0.05 and 50 iterations, the contrast differentiation parameter is c, and the minimal size of a recognized feature is f pixels. For comparison, a direct segmentation of the images has been performed as well, applying Steps 3 and 4 of Algorithm 1 with the same values of c and f directly to the image data. Together with the automatic counts, their accuracy is given in terms of the relative error (in percents) related to the manual counts from the third column.

ISSN: 1471-2415