High Spatiotemporal Precision Mapping of Optical Nanosensor Array Using Machine Learning

Optical nanosensors, including single-walled carbon nanotubes (SWCNTs), provide real-time spatiotemporal reporting at the single-molecule level within a nanometer-scale area. However, their superior sensitivity also makes them susceptible to slight environmental influences such as reference analytes in media, external fluid flow, and mechanical modulations. Consequently, they often fail to achieve the optimal limit of detection (LOD) and frequently convey misinformation spatiotemporally. To address this challenge, we developed a single-pixel mapping technique for optical nanosensor arrays that operates with high spatiotemporal precision using machine learning. We systematically measured the spatial sensing images of various analyte concentrations below the LOD by using a near-infrared (nIR) fluorescent SWCNT nanosensor array. For dopamine (DA) as an example analyte, we extracted single-pixel level sensing features such as entropy, the Laplacian operator, and neighboring values under noise levels. We then trained the artificial intelligence (AI) model to accurately identify specific reaction pixels of the nanosensor array, even below the LOD region. Additionally, our method can distinguish subtle noise caused by fluid in the media or mechanical modulation of the array substrate. As a result, our approach significantly improved the detection sensitivity of the nanosensor array, achieving a 13-fold increase over the original LOD and halving the detection time of the reporter pixels, with F1 scores exceeding 0.9. This method not only lowers the LOD of optical nanosensors but also isolates sensor responses specific to the analyte, providing accurate spatiotemporal information to the user, even in noisy conditions. It can be universally applied to various optical nanosensor materials and analytes, maximizing the sensitivity and accuracy of the nanosensors used in diagnostics and analysis.