Validation of a 3D printed smartphone-based lensmeter enhanced by AI for lens power measurements

Purpose : Conventional manual lensmeters are expensive and often too bulky for use in vision screenings. We recently developed a 3D printed smartphone-based lensmeter to measure ophthalmic lens power. Our findings showed that lens powers determined manually using this lensmeter were highly comparable to those obtained with a digital auto-lensmeter. Based on these findings, this project aimed to determine the accuracy and performance of using Artificial Intelligence (AI) for lens power determination.

Methods : A 3D printed lensmeter was developed to mount on a smartphone. A target with concentric rings formed by dots was captured by its camera. An AI algorithm, which processed the captured images using image pre-processing and k-means clustering, established the relationship between image dimensions and lens power (-10.00 to +5.00 D in 0.25 D steps) using stock lenses of refractive index (n) of 1.53. The measured values and the ground truth labels were used to fit a regression model that predicts lens power. The validity of the smartphone lensmeter was then evaluated with spherical lenses of varying refractive indices (n = 1.50, 1.56, and 1.67 in 1 D steps), plano-cylindrical lenses (-1.00 to -5.00 D, in 1 D steps, n = 1.53), and sphero-cylindrical lenses (n = 1.50).

Results : Our results showed that the smartphone lensmeter achieved high correlations for measuring the powers of spherical, plano-cylindrical, and sphero-cylindrical lenses (R = 1.00, p < 0.05) when compared with the digital auto-lensmeter. No statistically significant differences were found for spherical lenses (p = 0.74, t = 1.97), plano-cylindrical lenses (p = 0.84, t = 2.00), and sphero-cylindrical lenses (p = 0.90, t = 2.00). Mean absolute errors (MAE) for spherical, plano-cylindrical, and sphero-cylindrical lenses were 0.20 ± 0.16 D, 0.12 ± 0.08 D, and 0.22 ± 0.24 D, respectively. Measurement errors within ±0.25 D were found to be 68% for spherical lenses, 100% for plano-cylindrical lenses, and 81% for sphero-cylindrical lenses. Bland-Altman plots showed no consistent bias and good agreement for all lenses.

Conclusions : Our results strongly suggest that the AI-enhanced smartphone-based lensmeter offers accurate and efficient evaluation of spherical and cylindrical lens powers. This 3D printed smartphone lensmeter is a promising tool for application in clinics and vision screenings, especially in underprivileged and rural areas.

This abstract was presented at the 2025 ARVO Annual Meeting, held in Salt Lake City, Utah, May 4-8, 2025