Deep-learning enabled label-free bio-aerosol sensing using mobile microscopy (Conference Presentation)

Abstract

There is an increasing but unmet need for accurate, label-free and automated bio-aerosol sensing. To address this need, we developed a high-throughput, cost-effective and portable bio-aerosol sensor based on computational microscopy and deep-learning. Our device is composed of an impactor and a lens-less digital holographic on-chip microscope. It screens air at 13 liters per minute, and captures bio-aerosols on the impactor substrate. An image sensor then records the in-line holograms of these captured bio-aerosols in real time. Using these recorded in-line holograms, the captured bio-aerosols are analyzed within a minute, facilitated by two deep convolutional neural networks (CNNs): the first CNN simultaneously performs auto-focusing and phase-recovery to reconstruct both the amplitude and phase images of each bio-aerosol with sub-micron resolution; and the second CNN performs automatic classification of the reconstructed bio-aerosols into pre-trained classes and counting their densities in air. As a proof-of-concept, we demonstrated reconstruction and label-free sensing of five different types of bio-aerosols: Bermuda grass pollen, oak tree pollen, ragweed pollen, Aspergillus spore, and Alternaria spore. These bio-aerosols form some of the most common allergens in air. Using our mobile bio-aerosol sensor, we achieved ~94% precision and recall in differentiating these bio-aerosols without the use of any labeling. We also demonstrated successful sensing of oak tree pollens in the field using our mobile device. To the best of our knowledge, this is the first demonstration of automated label-free sensing of bio-aerosols using a portable device, which is enabled by computational microscopy and deep-learning. It has broad applications in label-free bio-aerosol sensing and air-quality monitoring.