The impact of different thresholds on optical coherence tomography angiography images binarization and quantitative metrics

Abstract

AbstractOptical coherence tomography angiography (OCTA) provides several data regarding the status of retinal capillaries. This information can be further enlarged by employing quantitative metrics, such as vessel density (VD). A mandatory preliminary step of OCTA quantification is image binarization, a procedure used to highlight retinal capillaries on empty background. Although several binarization thresholds exist, no consensus is reached about the thresholding technique to be used. In this study, we tested thirteen binarization thresholds on a dataset made by thirty volunteers. The aim was to assess the impact of binarization techniques on: (I) detection of retinal capillaries, assessed by the calculation of overlapping percentages between binarized and original images; (II) quantitative OCTA metrics, including VD, vessel tortuosity (VT) and vessel dispersion (Vdisp); (III) foveal avascular zone (FAZ) detection. Our findings showed Huang, Li, Mean and Percentile as highly reliable binarization thresholds (p < 0.05), whereas the worst binarization thresholds were Intermodes, MaxEntropy, RenylEntropy and Yen (p < 0.05). All the thresholds variably underestimated VD metric and FAZ detection, with respect to the original OCTA images, whereas VT and Vdisp turned out to be more stable. The usage of a Fixed threshold resulted extremely useful to reduce VD and FAZ underestimations, although bound to operators’ experience.