Deep learning-enhanced image analysis for liquid crystal optical sensing

In liquid crystal (LC) sensors, each microliter of LC contains billions of molecules with numerous orientation combinations, generating thousands of optical images with diverse textures under polarized optical microscope according to internal or external actuation. In this work, we utilize the VGG16 (Visual Geometry Group) deep learning (DL) model to accelerate the analysis of LC optical images, enabling visualized and precise sensing applications. The trained model helps improve the LC sensing speed and sensitivity to achieve a classification accuracy of 0.9113 within 30 s when triggered by two representative surfactants, cetyltrimethylammonium bromide (CTAB) and sodium dodecyl sulfate (SDS). The average relative errors are reduced to 3.54 and 7.94%, respectively, in quantitatively sensing the insulin-specific aptamer and insulin. In addition, the sensing time decreases from 300 s (using gray scale intensity quantification) to 90 s for insulin recognition and concentration detection. DL has been proven to be a useful and powerful analytical tool in image analysis, improving the speed and accuracy of optical image-based sensors.