Machine Learning-Based Optimization of Lens Design for Capsule Endoscopes

Abstract

This study presents a novel optimization framework for capsule endoscope (CE) lens design, integrating optical ray-tracing simulations with probabilistic machine learning-based model. Traditional optical lens design approaches, relying predominantly on local optimization techniques available in ray-tracing software such as Zemax, often encounter challenges related to local minima and inefficiencies in exploring extensive parameter spaces. To address these limitations, this research adopts a systematic approach combining LHS and Machine Learning models. Based on existing CE lens designs documented in literature a modified design is simulated in Zemax for white light imaging. LHS was used to generate the design points of critical optical parameters to fill the design space efficiently. Each generated candidate LHS design is evaluated using Zemax simulations to calculate the single objective metric, RMS spot radius. This RMS spot value, representing the average performance across five distinct fields, quantifies the optical system's imaging capability. Machine learning models, trained on these data, enable rapid and accurate prediction of RMS spot radius across the entire parameter space, significantly enhancing optimization. Results demonstrate the effectiveness and efficiency of this machine learning model-based single-objective optimization method, providing an optimized lens design configuration with substantially improved imaging performance compared to traditional optimization techniques.