Multiple-object geometric deformable model for segmentation of macular OCT: errata

Abstract

Boundary errors were incorrectly computed in our paper [ Biomed. Opt. Express 6( 4), 1063 ( 2014)], which resulted from the manual segmentations being incorrectly converted between file formats. In particular, our paper mischaracterized the error of the method in comparison to that of Lang et al. [ Biomed. Opt. Express 6( 4), 1133– 1152 ( 2013)]. We include corrected tables, replacing Tables 1 and 2 in [ Biomed. Opt. Express 6( 4), 1063 ( 2014)].

1. Introduction

The reported errors in [1] were mistakenly computed, due to an an error in conversion between file formats. We include corrected versions of Tables 1 and 2.

Table 1.

Mean (and standard deviations) of the Dice Coefficient across the eight retinal layers. A paired Wilcoxon rank sum test was used to test the significance of any improvement between RF+Graph [2] and our method, with strong significance (an α level of 0.001) in two of the eight layers. However, RF+Graph was also significantly better than MGDM on two of the eight layers. In three of the four remaining layers, MGDM is marginally better than RF+Graph without reaching statistical significance.

Layer	Dice Coefficient		P-Value
	RF+Graph [2]	MGDM [1]
RNFL	0.9194 (±0.0378)	0.9205 (±0.0239)	0.976
GCIP	0.9344 (±0.0385)	0.9267 (±0.0276)	< 0.001 †
INL	0.8638 (±0.0249)	0.8437 (±0.0293)	< 0.001 †
OPL	0.8786 (±0.0159)	0.8814 (±0.0194)	0.118
ONL	0.9391 (±0.0118)	0.9414 (±0.0127)	0.208
IS	0.8560 (±0.0363)	0.8558 (±0.0202)	0.940
OS	0.8467 (±0.0442)	0.8675 (±0.0317)	< 0.001 ★
RPE	0.8931 (±0.0438)	0.9119 (±0.0312)	< 0.001 ★

^★Indicates that MGDM was significantly better than RF+Graph.

^†Indicates that RF+Graph was significantly better than MGDM.

Table 2.

Mean absolute errors (and standard deviation) in microns for our method (MGDM) in comparison to RF+Graph [2] on the nine estimated boundaries. A paired Wilcoxon rank sum test was used to compute p-values between the two methods with strong significance (an α level of 0.001) in six of the nine boundaries. However RF+Graph was significantly better than MGDM on one of the nine boundaries.

Boundary	Absolute Errors		P-Value
	RF+Graph [2]	MGDM [1]
ILM	2.695 (±0.3591)	2.379 (±0.3990)	< 0.001 ★
RNFL-GCL	4.317 (±2.4034)	4.016 (±1.8566)	< 0.001 ★
IPL-INL	3.935 (±0.6803)	4.219 (±1.4218)	0.644
INL-OPL	3.467 (±0.6546)	4.237 (±1.1758)	< 0.001 †
OPL-ONL	3.230 (±0.8003)	3.200 (±0.9179)	0.386
ELM	3.033 (±0.8060)	2.675 (±0.5502)	< 0.001 ★
IS-OS	2.295 (±0.6965)	1.928 (±0.7384)	< 0.001 ★
OS-RPE	4.251 (±1.6164)	3.965 (±1.6033)	< 0.001 ★
BrM	3.103 (±1.7818)	2.856 (±1.7498)	< 0.001 ★
Overall	3.370 (±1.4080)	3.275 (±1.4909)	< 0.001 ★

^★Indicates that MGDM was significantly better than RF+Graph.

^†Indicates that RF+Graph was significantly better than MGDM.

2. Discussion and conclusion

Our method (MGDM) still significantly outperforms RF+Graph in six of the nine boundaries and has better overall accuracy. The errors reported for both methods are lower than those reported in [1] and the mean boundary errors reported for the RF+Graph method are now consistent with those reported in [2]. Note that the within subject variation was not included in our computation of the standard deviation, which is why our standard deviations are significantly lower than those reported in [2].