Deep learning enables accurate analysis of images generated from droplet-based digital polymerase chain reaction (dPCR)

Accurate analysis of digital polymerase chain reaction (dPCR) images is a key to the successful miniaturization of a dPCR system, which would allow point-of-care testing, but has been hampered by several issues, mainly out-of -focus images, and variations in shape, size, and fluorescence intensities of the droplets. Herein, we performed semantic segmentation by using a deep learning model combined with the circle Hough transform (CHT). For the deep learning model, Attention DeepLabV3+ was modified and used to analyze positive and negative droplets in the low-quality dPCR images, while CHT helps to count the droplets. The modified Attention DeepLabV3+ was trained with augmented 4224 dPCR images, which were sufficient for deep learning model training. As a proof of concept, encapsulated genomic DNA (gDNA) of Escherichia coli O157:H7, a foodborne pathogen, was amplified, and the corresponding droplets were analyzed with the deep learning model to demonstrate its rapid, absolute quantification and accurate identification of the target gDNA. Our analysis method achieved similar accuracy to the commercial, bulky dPCR system, which is based on continuous-flow analysis methods. Furthermore, the deep learning-based approach exhibited a great sensitivity for detecting pathogenic bacteria from low-quality images with a low detection limit of 4.0 x 101 copies/mu L. Compared with the commercially available instrument, the deep learning method demonstrated a similar molecular analysis performance. The presented results may lay the ground for efficient absolute quantification of target nucleic acids from the dPCR images as well as the estab-lishment of a point-of-care analysis platform.