Figure 2 | Scientific Reports

Figure 2

From: Unbiased identification of novel subclinical imaging biomarkers using unsupervised deep learning

Figure 2

Representative examples of feature maps obtained by the local embedding. The composites to the right of each column show heatmaps of conventional biomarkers obtained by validated automated image segmentation algorithms11, 12. High and low activation of the detected new biomarkers with concomitant visual function are shown side-by-side. Top row: Feature (a5) demonstrates a pronounced negative structure-function correlation, despite a low correspondence to retinal fluid, which is the conventional marker attributed a high relevance for vision. We assume that this biomarker candidate corresponds to subretinal hyperreflective material (arrow). Middle row: Feature (a17) demonstrates the best correlation with markers of exudation as conventionally measured in OCT. An excellent correspondence is for instance observed for intraretinal cystoid fluid (compare the lobulated pattern). Bottom row: Feature (a4) represents a new subclinical biomarker candidate discovered in this work (arrows). The marker does not intrinsically correspond to previously reported clinical entities in OCT images. Remarkably, a positive correlation between the activation of a4 and visual function markers was noted. Color bars indicate the activation level from maxiumum (dark) to minimum (light). IRC, intraretinal cystoid fluid; PED, pigment epithelial detachment; RT, retinal thickness; SRF, subretinal fluid; SHRM, subretinal hyperreflective material.

Back to article page