In the current study, we used dimensionality reduction using UMAP and subsequent clustering for automatically detected NPAs on widefield OCTA images in DR. This approach elucidated several patterns of NPAs distribution in DR, namely; minimal type, mild type, temporal type, inferotemporal type, both superior and inferior types, and macular type. Comparisons of the NPA extents among clusters encouraged us to infer the unique progression of NPAs in this cross-sectional study, which should be confirmed by future longitudinal studies. Further investigation demonstrated that the patterns of NPA distribution are associated with DR severity to some extent, although the amounts of NPAs did not always correspond to DR severity.
The NPS ratios were the lowest around the optic disc and along the superotemporal and inferotemporal vascular trunks in all 201 eyes. This observation may be explained by an additional capillary layer, the radial peripapillary capillaries, and higher perfusion pressure there.
16,28 Furthermore, NPAs were more frequently seen in the extramacular areas than in the macula. This may depend on the predisposition of vascular structure specific to the extramacular and macular regions.
18 Because the temporal raphe is the most distal to the optic disc, perfusion pressure is the lowest and capillary nonperfusion might be more likely to progress in the temporal subfield.
20,29
Eyes in cluster 1 had minimal NPAs, although some eyes suffered from PDR. We were able to evaluate vascular lesions from the macula to the midperipheral regions using widefield OCTA images, although peripheral images could not be obtained. Recent publications showed that some cases have peripheral lesions including NPAs on ultrawide field FA images.
4,12,13 Future studies should elucidate the status of capillary nonperfusion in the periphery on ultrawide field montage of OCTA images. Empirically, some eyes of young patients with PDR had no NPAs, which might not be compatible with the general concept that VEGF derived from NPAs develops neovascularization.
Clusters 3 (temporal type) and 5 (both superior and inferior types) showed an exclusive relationship with respect to the NPA distribution. We therefore hypothesized that blood flow redirects toward the remaining patent vessels after the NPA extension in the specific areas. When NPAs extended in the temporal subfield, blood flow was maintained in the superior and inferior peripheral regions in eyes of cluster 3. We might explain that the reduced amounts of capillaries in the temporal subfield increase perfusion pressure and concomitant resistance to capillary obstruction in other retinal regions. In eyes of cluster 5, segmental perfusion in the superonasal and inferonasal subfield might contribute to NPA progression there. As a result, the perfusion pressure might increase in the temporal subfield, and the seamless network in the deep capillaries there might also make NPAs progression less likely.
18
In eyes of cluster 4, the NPA in the inferotemporal region could be attributed to 2 factors: the trait during the development and the perfusion pressure. First, the retina is thinner and vascular density is lower in the inferior subfield than in the superior subfield in healthy adults.
30,31 This might lead to reduced vascular density or segmental perfusion of the extramacular regions in the inferior subfield. The second point is that lower perfusion pressure on the temporal side than on the nasal side makes it more likely for NPA to develop in the inferotemporal region. The comparative study may support the possible transition from clusters 4 to 5, although they were far from each other in the UMAP projection.
The heatmaps of clusters 1 and 6 seemed similar to each other, although there were statistical differences in NPS percentages of several sectors. The main difference was the NPS counts in the central 1 mm, and we speculated that eyes in cluster 6 might correspond to those with DMI.
3 The foveal avascular zone is the most distal to the optic disc and parafoveal capillaries are composed of only one or two layers, compared to three or more layers in other areas within the macula.
14,15,32 Additionally, deep capillaries may reduce in eyes with DME, and VEGF derived from retinal neurons might also decrease in eyes with diabetic neurodegeneration.
33–35 However, it should be noted that we cannot exclude the possibility of artifacts from the OCTA machine with an artificial intelligence (AI)-driven denoise algorithm or the false negative in the image processing of vessels with high density.
36,37
The diversity in the NPA distribution suggests that multiple factors influence the progression of capillary nonperfusion. In this cross-sectional study, we could not guarantee the processes of NPA progression completely, although the statistical comparisons suggested that there are eight pathways of the NPA progression. The transition from cluster 1 to 2 might correspond to the initiation; at random development of minimal NPAs. Because clusters 3 and 4 were near cluster 1 on the UMAP projection, the unknown and specific factors or predispositions might promote the NPAs extension to the specific regions vigorously. After the initiation, four pathways for NPA progression might lead to the specific distribution as discussed above. However, the probabilistic analyses suggest less frequency of NPA progression from cluster 2 to clusters 3, 4, 5, and 6. The probabilities of the transitions were higher than those of random transitions (data not shown). However, future longitudinal studies should confirm the pathways for NPA progression and the actual probability for the transition between clusters per year.
There are several limitations to this study. The inclusion and exclusion criteria were applied in this single-center study, which may result in selection bias. The OCTA machine used in this study delineates angiographic images from the macula to the midperiphery, but not to the periphery. The image quality may affect the automatic quantification. Because the longer time was required for image acquisition of wide-field OCTA, images of sufficient quality were not obtained in eyes with poor fixation and concomitantly there was another selection bias. The artifacts at OCTA image acquisition may modify the subsequent assessment.
38,39 We used images of superficial and deep layers in this study, and future analyses of layer-specific NPAs would provide more detailed and accurate data. Although image processing in this study had advantages in the NPA distribution, the vessel edges could not be detected completely, and the NPA amounts were approximately measured. The presence of NVE could potentially lead to an underestimation of the NPA. We selected a UMAP algorithm to perform dimensionality reduction, and future studies may compare the results from other algorithms. We tried to infer the NPA progression from the macula to the midperiphery using the cross-sectional study, although future longitudinal studies should be planned to confirm the characteristics of the NPA progression in the whole retinas and to elucidate whether the patterns of NPA progression explain the differences in DR severity or visual impairment.
40
In conclusion, the dimensionality reduction using the UMAP algorithm and subsequent clustering divided DR eyes into six clusters based on the NPA distribution in the widefield OCTA images. Statistical and probabilistic analyses allowed us to infer the pathways for the NPA progression.