Abstract
Purpose :
Use a machine learning approach to reduce optical aberrations and enhance visual acuity for B-KPro Type 1 recipients, focusing on correcting spherical aberration, inherent to the spherical geometry of the lens.
Methods :
Using Zemax OpticStudio, a B-KPro model was simulated, and an evolutionary algorithm, focusing on anterior surface, was employed to optimize spherical aberration coefficient. An evolutionary algorithm was used due to its specific efficiency with non-linear processes. The vectors were the coefficients of the geometric equation for the anterior surface of the KPro lens. The fitness function took into consideration the total optical aberrations of the system and the desired back focal length.
Results :
The machine learning approach led to a re-shaping of the optical profile of the B-KPro lens with reduced aberrations from 0.09725 to 0.043827. Simulated visual quality improvement was evident in an extended scene analysis (Fig. 1). The Modulation Transfer Function (MTF) curve demonstrated notable reduction in spherical aberration, which enhanced contrast sensitivity at higher spatial frequencies (Fig. 2). An ablation study highlighted the importance of the conic and 2nd order coefficients, while the 4th order and all subsequent coefficients showed minimal impact on image quality.
Fig. 1: The geometric image analysis from model prior to optimization (i), after optimizing for the conic coefficient (ii), after changing the conic and 2nd order coefficients (iii) and after changing the conic, 2nd, 4th and so on to the 16th order coefficient respectively (iv).
Fig. 2: The images are MTF curves before (i) and after (ii) introducing asphericity with the algorithm showing a total area under the curve of 0.2287 and 0.3646 respectively and a cutoff frequency of 108.2 and 123 respectively.
Conclusions :
The evolutionary algorithm successfully identified key variables to reduce aberrations in the B-KPro system. While this simplified simulation has inherent assumptions, its analytic power and Zemax-integrated machine learning make it a robust lens optimization tool. Future iterations could explore more comprehensive frameworks, like neural networks, to optimize additional variables and this work may be applicable to other optical lenses using in ophthalmology. Further evaluation of these results will require lens manufacturing and evaluation in a simulated eye, as shown previously by our group.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.