Exploring the association of collaterals and vessel density using optical coherence tomography angiography in retinal vein occlusions

Purpose Using optical coherence tomography angiography (OCTA) to characterize the types of collaterals in eyes with retinal vein occlusion (RVO) and further investigate their correlations with vessel densities of the superficial (SCP) and the deep capillary plexus (DCP). Methods This cross-sectional study included 25 eyes of 23 patients with RVO. 3 × 3 mm2 OCTA macular scans were used to quantify the parafoveal vessel density (VD) of the SCP and DCP, and to classify the collaterals into one of four types (true superficial, true deep, superficial diving, and foveal collateral). Poisson regression model was used to identify significant associations between parafoveal VD and collaterals. We further compared parafoveal VD between subgroups classified by the presence of specific collateral types based on the results of a clustering algorithm. Results 16 of 25 eyes (64%) developed collaterals. Of the 43 collateral vessels analyzed, 12/19 (63%) true superficial collaterals developed in eyes with central RVO, while all 10 superficial diving collaterals (100%) developed in eyes with branch RVO. Located exclusively in the SCP, true superficial collaterals were all arteriovenous (A-V), while diving collaterals were all veno-venular (V-V). We found a significant negative correlation between SCP VD and the total number of collaterals (R2 = 0.648, P < 0.001) for the entire study cohort. Furthermore, BRVO eyes that developed superficial diving collaterals and CRVO eyes that developed true superficial collaterals demonstrated statistically significant decrease in SCP VD (P-value = 0.014) and DCP VD (P-value = 0.030), respectively, as compared to their counterparts. Conclusion Our data shows that decreased capillary perfusion in RVO is associated with the development of collaterals, while the RVO type largely dictates the type of collateral that ultimately develops.

RTVue-XR Avanti system (Optovue Inc., Fremont, California, USA) was used to 120 acquire a 3 X 3 mm 2 scanning area, centered on the fovea.

155
All collateral vessels were identified within the 3 X 3 mm² en face angiograms of 156 each eye. Images were further analyzed to assess the location of each collateral vessel 157 using the corresponding cross-sectional scans with angiographic flow overlay, and 3D 158 Projection artifact removal (PAR) technology to eliminate projection artifact. We   (Table 1) into 208 distinct groups using the selected features: "RVO Type", "LogMAR BCVA", "SCP VD" 209 and "DCP VD". Using this algorithm, it is expected that similar patterns will be observed 210 within the clusters and different patterns between the clusters. Since "RVO Type" in 211

222
(1) Randomly choose k data records as the medoids

223
(2) For every data record, find and group with its closest medoid

224
(3) For each cluster, find the data record with the lowest average distance to the 225 rest of the data records within the cluster. If the data record is different from the previous medoid, replace the previous medoid with the data record as the new 227 medoid.
228 Any changes to the medoids in this procedure will cause the algorithm to repeat Step 2.
229 If all the medoids remain unchanged, the algorithm will end the procedure. The mean parafoveal VD (%) in the SCP and DCP within a 3 x 3-mm 2 area were 255 39.8 ± 5.6 and 42.9 ± 5.6, respectively. SCP VD or DCP VD values below one standard 256 deviation of their corresponding mean values were considered "low". 43 collateral 257 vessels were identified and analyzed (mean, 1.7; range, 0-5 collaterals per eye). We 258 found that the number of collaterals was highest ( 4 collaterals per eye) in eyes with 259 either low SCP or DCP VD (Table 3).

260
We examined the overall mean VD of the RVO types, classified by presence of 261 collaterals, as shown in    306 BRVO, and HRVO), rather than vessel density. Clustering revealed a similarity in 307 collateral types within each of these groups, with true superficial collaterals (Fig 1) 308 occurring predominantly in the CRVO and HRVO groups, while the superficial diving 309 collaterals (Fig 2) were found in the BRVO group. Furthermore, multivariate analysis of 310 the entire study cohort showed that non-perfusion at the SCP capillary plexus in RVO 311 was most significantly associated with the total number of collaterals (P < 0.001). The 312 apparent lack of similar association of DCP VD could be due to the heterogeneity of our 313 study patients including the type of occlusion present (BRVO, CRVO, and HRVO) [29] 314 that may have introduced variations to the regression model (S1 Table). It is possible 315 that with a greater number of cases, a stronger correlation between DCP VD and the 316 overall number of collaterals in RVO could be revealed.

317
When we divided the eyes into RVO subgroups based on the aforementioned 318 clustering results, BRVO eyes that developed superficial diving collaterals had 319 significantly lower SCP VD (P = 0.014, Mann-Whitney U test) than those that did not. It 320 is possible that diving collaterals form as a secondary complication in response to 321 primary SCP capillary closure. In this model, diving collaterals (Fig 2A-2D) would serve 322 as a channel for SCP venous outflow, bypassing the occluded SCP capillary bed, with 323 the DCP providing a more convenient outflow through its highly connected vortex Strengths of our study include the stringent criteria for quality of our datasets, 370 including high-signal strength index, 3 x 3-mm 2 OCTA scans which made it possible to 371 obtain the most accurate of VD in the macula at the different plexuses [42,43]. Due to 372 the limitation of the scan size, however, our analysis did not include the peripheral 373 collaterals. In addition, we carefully assessed the OCTA cross-sectional scans to trace 374 the flow signals of the collateral vessels in the en face images (Fig 1 and 2) and