Skip to main content
Fig. 2 | International Journal of Retina and Vitreous

Fig. 2

From: Comparing code-free deep learning models to expert-designed models for detecting retinal diseases from optical coherence tomography

Fig. 2

Saliency maps (XRAI method) of accurately predicted cases by the CFDL image model on the left and the best bespoke image (Xception) model on the right. Overall, both models highlighted similar areas as the most important region on OCT for the predicted class. (A) Normal macula: both saliency maps highlight the hyperreflective outer retinal layers, corresponding to the retinal pigmental epithelium/Bruch’s membrane complex and interdigitation zone. (B) Epiretinal membrane: the CFDL map highlights the epiretinal membrane/ internal limiting membrane while the bespoke map focuses on the overall foveal architecture. (C) Macular hole: both maps highlight the anvil-shaped deformity of the edges of the retina and the intraretinal edema. (D) Diabetic macular edema: both maps highlight intraretinal cystoid spaces. (E) Wet age-related macular degeneration: both maps highlight the elevation of the RPE. The CFDL map also highlights a small pocket of subretinal fluid

Back to article page