PURPOSE: To compare area measurements between swept source optical coherence tomography angiography (SSOCTA), fluorescein angiography (FA), and indocyanine green angiography (ICGA) after applying a novel deep-learning-assisted algorithm for accurate image registration. METHODS: We applied an algorithm for the segmentation of blood vessels in FA, ICGA, and SSOCTA images of 24 eyes with treatment-naive neovascular age-related macular degeneration. We trained a model based on U-Net and Mask R-CNN for each imaging modality using vessel annotations and junctions to estimate scaling, translation, and rotation. For fine-tuning of the registration, vessels and the elastix framework were used. Area, perimeter, and circularity measurements were performed manually using ImageJ. RESULTS: Choroidal neovascularization lesion size, perimeter, and circularity delineations showed no significant difference between SSOCTA and ICGA (all P > 0.05). Choroidal neovascularization area showed excellent correlation between SSOCTA and ICGA (r = 0.992) and a Bland-Altman bias of -0.10 ± 0.24 mm. There was no significant difference in foveal avascular zone size between SSOCTA and FA (P = 0.96) and an extremely small bias of 0.0004 ± 0.04 mm and excellent correlation (r = 0.933). Foveal avascular zone perimeter was not significantly different, but foveal avascular zone circularity was significantly different (P = 0.047), indicating that some small cavities or gaps may be missed leading to higher circularity values representing a more round-shaped foveal avascular zone in FA. CONCLUSION: We found no statistically significant differences between SSOCTA and FA and ICGA area measurements in patients with treatment-naive neovascular age-related macular degeneration after applying a deep-learning-assisted approach for image registration. These findings encourage a paradigm shift to using SSOCTA as a first-line diagnostic tool in neovascular age-related macular degeneration.
PURPOSE: To compare area measurements between swept source optical coherence tomography angiography (SSOCTA), fluorescein angiography (FA), and indocyanine green angiography (ICGA) after applying a novel deep-learning-assisted algorithm for accurate image registration. METHODS: We applied an algorithm for the segmentation of blood vessels in FA, ICGA, and SSOCTA images of 24 eyes with treatment-naive neovascular age-related macular degeneration. We trained a model based on U-Net and Mask R-CNN for each imaging modality using vessel annotations and junctions to estimate scaling, translation, and rotation. For fine-tuning of the registration, vessels and the elastix framework were used. Area, perimeter, and circularity measurements were performed manually using ImageJ. RESULTS: Choroidal neovascularization lesion size, perimeter, and circularity delineations showed no significant difference between SSOCTA and ICGA (all P > 0.05). Choroidal neovascularization area showed excellent correlation between SSOCTA and ICGA (r = 0.992) and a Bland-Altman bias of -0.10 ± 0.24 mm. There was no significant difference in foveal avascular zone size between SSOCTA and FA (P = 0.96) and an extremely small bias of 0.0004 ± 0.04 mm and excellent correlation (r = 0.933). Foveal avascular zone perimeter was not significantly different, but foveal avascular zone circularity was significantly different (P = 0.047), indicating that some small cavities or gaps may be missed leading to higher circularity values representing a more round-shaped foveal avascular zone in FA. CONCLUSION: We found no statistically significant differences between SSOCTA and FA and ICGA area measurements in patients with treatment-naive neovascular age-related macular degeneration after applying a deep-learning-assisted approach for image registration. These findings encourage a paradigm shift to using SSOCTA as a first-line diagnostic tool in neovascular age-related macular degeneration.