Automated quantification of RETeval™ sensor electrode position from infrared eye images: a preliminary technical report.

PURPOSE: To present an automated method for computing the position of the RETeval skin sensor electrode and the iris color ratio (ICratio) from infrared eye images extracted from RETeval recording files. METHODS: Infrared eye images were extracted from .RFF binary files recorded in 19 healthy young participants (mean age 10.6 years, range 5.0-15.8; 7 male, 12 female) using the RETeval 6-step Light First Troland ISCEV protocol. A multilabel deep learning segmentation model (SegFormer) was trained to delineate six ocular structures: pupil, iris/limbus, corneal reflex, visible white bulbar area, electrode, and inner lens ring. Electrode position was quantified as the signed displacement vector from the corneal reflex centroid to the electrode arc-contact point (Pₑ), decomposed into horizontal (dx) and vertical (dy) components in millimeters. The iris color ratio (ICratio) was derived from the segmentation masks as the mean gray level of the iris annular ring divided by the median gray level of the dark pupil, with corneal reflex pixels excluded. RESULTS: The segmentation model achieved a mean validation Dice coefficient of 0.971 across the six classes on a held-out recording session. Infrared images (n = 271) were extracted from 19 participants (38 eyes); 67 (24.7%) satisfied all three quality-control criteria, and an electrode measurement was obtained for all 38 eyes. Vertical electrode displacement (dy) was 8.57 ± 1.81 mm (right eye) and 9.14 ± 1.89 mm (left eye; p = 0.13 between eyes); horizontal displacement (dx) was - 3.11 ± 2.42 mm and - 3.70 ± 4.74 mm, respectively (p = 0.67). Euclidean displacement (|d|) differed significantly between eyes (9.45 ± 1.63 vs 10.74 ± 2.59 mm; p = 0.004), while angular displacement did not (p = 0.81). Mask-based ICratio was 1.128 ± 0.036 (right) and 1.121 ± 0.039 (left), with a significant interocular correlation (r = 0.811, p = 0.008); the device-derived ratio was 1.134 ± 0.024 and 1.149 ± 0.038 (interocular r = 0.561, p = 0.12). The two methods differed by - 0.016 on average (95% limits of agreement - 0.108 to + 0.077; p = 0.14). CONCLUSIONS: This preliminary study provides an automated method for determining the position of the RETeval sensor electrode from the infrared images recorded by the device. A measure of iris darkness derived from the same segmentation is reported as a secondary technical output.