Deep Learning-Based Age-Related Macular Degeneration Detection Using Combined OCT and Fundus Images

Abstract

In order to identify age-related macular degeneration (AMD), researchers have recently developed new deep learning (DL) models that use a single visual modality. The most crucial modalities for examining AMD in clinical settings are retinal fundus and optical coherence tomography (OCT) images. It is unclear whether using fundus and OCT data simultaneously in the DL approach is advantageous. OCT and fundus imaging data from postmortems from Project Macula were used in this experimental investigation. To diagnose AMD, the DL based on OCT, fundus, and a combination of OCT and fundus was developed.
Pre-trained VGG-19 and random forest transfer learning comprised these models. The DL utilizing OCT alone demonstrated diagnostic efficiency after data augmentation and training, with an accuracy rate of 82.6% (81.0–84.3%) and an area under the curve (AUC) of 0.906 (95% confidence interval, 0.891–0.921). The accuracy rate of the DL utilizing fundus alone was 83.5% (81.8–85.0%) and the AUC was 0.914 (0.900–0.928). When the fundus and OCT were used together, the diagnostic power rose with an accuracy rate of 90.5% (89.2–91.8%) and an AUC of 0.969 (0.956–0.979). The DL using both OCT and fundus data performed better than the DL using OCT alone (P value < 0.001) and fundus image alone (P value < 0.001), according to the Delong test. Compared to deep belief network techniques (P value = 0.042) and a constrained Boltzmann machine (P value = 0.002), this multimodal random forest model performed even better. Duncan's multiple range test revealed that the multimodal approaches far outperformed the single-modal approaches. When compared to this data alone, a multimodal DL algorithm based on the combination of OCT and fundus image improved the diagnostic accuracy in this pilot investigation. For a more accurate diagnosis of AMD, future diagnostic DL must use the multimodal approach to integrate different imaging modalities.