Purpose
To predict future visual acuity (VA) during the patient-specific anti-VEGF treat-and-extend phase, using a set of OCT images acquired during induction in treatment-naive patients with exudative age-related macular degeneration (AMD).
Methods
During induction with ranibizumab or bevacizumab, Topcon SS-OCT images were acquired every 2 weeks (Figure 1a), while during treat-and-extend, patients were only imaged on the day of injection Quantitative features were extracted, describing the underlying retinal structure, based on automated segmentation of layers and fluid regions. These include regional inner retina, outer nuclear layer (ONL), photoreceptor outer segment with retinal pigment epithelium layer, and total retinal thicknesses as well as regional intra-, sub-retinal fluid, and pigment epithelial detachment volume and area. An Early Treatment of Diabetic Retinopathy (ETDRS) grid centered on the fovea was used to compute these features for all 9 regions (Figure 1b). Random forest regression was used to predict the patient’s logMAR visual acuity two visits (visit 9) after the induction phase using leave one out validation.
Results
32 subjects were included in the study, average age was 78 years, 50% were female. Average interval from the end of induction to visit 9 was 12 weeks. Correlation coefficient of measured logMAR VA to predicted VA was R = 0.57, while bias and standard deviation were 0.06 and 0.16 logMAR, respectively (Figure 2). The two most important OCT-derived features were found to be mean ONL thickness in the inferior parafoveal region at visit 2 (2 weeks) and intraretinal fluid area in the superior parafoveal region at visit 6 (10 weeks from start of treatment).
Conclusions
We proposed and tested a methodology to predict the anti-VEGF functional response from a longitudinal series of OCT scans during the induction phase. The results of this pilot study are promising for our long term goal of image-guided prediction of visual outcome and treatment intervals for anti-VEGF treatments for CNV.