Evaluating GAN-Based Lens-less Digital In-line Holographic Microscope for Common Pollen Classification

Lens-less digital in-line holographic microscopy (DIHM) is a low-cost, wide-field imaging technique that computationally reconstructs focused and artifact-free images. Recently, deep learning methods have been applied for this reconstruction, with generative adversarial networks (GANs) showing particular promise. In this study, we investigated whether GAN-reconstructed DIHM images can be utilized for visual classification of common pollen types. Four pollens, Bermuda grass, Silver birch, Olive tree, and Corn, were imaged using both a brightfield slide scanner and a custom-built lens-less DIHM (658 nm laser diode, 2×2 µm pixel camera). Three veterinary cytopathologists classified pollen images. Brightfield microscopy achieved 97.9% accuracy, while DIHM reached 72.9%. DIHM also achieved around 90% accuracy (i.e., comparable to current state-of-the-art) either when birch and olive pollen were grouped or when considering the best-performing evaluator. These results suggest that GAN-based DIHM is a cost-effective alternative to conventional optical microscopy. Visual pollen classification could be further improved by optimizing GAN-based reconstruction methods, expanding training datasets, and enhancing evaluator training.