Vascular network of the human macula from OCT

Abstract

AbstractPurpose To compute the vascular network of the human macula from spectral‐domain optical coherence tomography (OCT) to an extent similar to that of color fundus photography (CFP).Methods Macular cube protocol scans of 512x128x1024 and 200x200x1024 voxels of 20 eyes from 13 type 2 diabetic patients and 10 eyes from 10 healthy volunteers were collected from the Cirrus HD‐OCT (Carl Zeiss Meditec, Dublin, CA, USA) database. Additionally, CFPs and fluorescein angiograms (FAs) from the patients' eyes were gathered from the imaging database. Three distinct fundus references were computed from the OCT volumes after proper preprocessing. An additional OCT fundus image (OCTref) is computed as the principal component of these 3 OCT fundus references. The visible vascular network was manually segmented on CFP, FA and OCTref for comparison. Finally, a support vector machine (SVM) pattern classification algorithm was used to classify each pixel of the OCTref image into the vessel or non‐vessel classes from a set of 14 features computed from the OCT fundus references.Results Over 67% (67.8(7.2)%, average(SD)) of the vascular network manually segmented from the FA was manually segmented from the CFP, while this percentage raises to 69.2(8.9)% for OCT. In this way, the OCTref allows to compute an extended vascular network as compared to the CFP (102.8(14.4)%). When comparing the automatic versus the manual vascular segmentation, a specificity of 99.4(0.2)% and a sensitivity of 83.9(4.0)% were obtained. Overall, the accuracy of the automatic classification is of 98.0(0.4)%.Conclusion The proposed algorithm allows for the segmentation of the vascular network from OCT scans of the ocular fundus to a level similar to that of color fundus photography.