Label-free Ambient Pollen Identification and Classification Using Total Internal Reflection Microscopy and Deep Learning

Ambient pollen identification and classification using traditional label-dependent methodologies is often time-consuming and prone to errors. Conventional methods such as bright field (BF) microscopy and including labeled techniques provide low-contrast and low signal-to-noise ratio images of pollens.  These techniques are also associated with high noise due to the complex nature of ambient samples.  Numerous other label-free techniques such as phase contrast, and quantitative phase imaging have been applied for ambient pollen imaging; however, these imaging methods provide better contrast images but the identification of pollens in ambient samples is very difficult due to the heterogeneous nature of ambient particles as it contains particulate matter, fungal spores, mold spores, and other ambient particles. To address these limitations, the current work employs a label-free novel application of total internal reflection (TIR) phenomenon. When the light beam undergoes TIR at an optical interface, an evanescent field is generated over the interface where the samples are prepared. The generated evanescent field illuminates the sample upto a certain depth (~500nm) only, which helps to avoid the background noise and gives high-contrast images with high SNR. TIR imaging enhances the optical properties of ambient pollen by emphasising surface and near-surface features, providing remarkable contrast in ambient pollen images. The study used TIR imaging, enabling non-invasive, no-sample preparation and precise identification of pollen in ambient samples, even with high background concentration of ambient particles.  It also allows better visualization of pollen boundary and additional surface features such as the polarity and aperture patterns. Additionally, the CNN-based deep-learning model is used for pollen detection, significantly advancing ambient pollen analysis. The model demonstrates strong performance, with a high F1 score for detecting pollen (0.83) and a well-balanced overall performance (F1 score of 0.77 for all classes). The confusion matrix shows excellent classification accuracy, especially for the pollen class. The model’s mean average performance is 76.7% across all classes at a threshold of 0.5, indicating good performance. Preliminary results demonstrate the model's robust performance, even when handling complex ambient samples with high ambient concentrations of other ambient particles. Pollen monitoring is crucial due to the scarcity of comprehensive data on airborne pollen, which impacts public health. The application of TIR microscopy combined with automated analysis offers a label-free, real-time, and field-deployable solution for addressing challenges in airborne particle monitoring. These results highlight the novel potential of TIR microscopy with deep learning as a method for precise, effective, and scalable pollen monitoring.