A deep learning based method for automatic analysis of high-throughput droplet digital PCR images

Abstract

Droplet digital PCR (ddPCR) is a technique for absolute quantification of nucleic acid molecules and is widely used in biomedical research and clinical diagnosis. ddPCR partitions the reaction solution containing target molecules into a large number of independent microdroplets for amplification and performs quantitative analysis of target molecules by calculating the proportion of positive droplets by the principle of Poisson distribution. Accurate recognition of positive droplets in ddPCR images is of great importance to guarantee the accuracy of target nucleic acid quantitative analysis. However, hand-designed operators are sensitive to interference and have disadvantages such as low contrast, uneven illumination, low sample copy number, and noise, and their accuracy and robustness still need to be improved. Herein, we developed a deep learning-based high-throughput ddPCR droplet detection framework for robust and accurate ddPCR image analysis, and the experimental results show that our method achieves excellent performance in the recognition of positive droplets (99.71%) within a limited time. By combining the Hough transform and a convolutional neural network (CNN), our novel method can automatically filter out invalid droplets that are difficult to be identified by local or global encoding methods and realize high-precision localization and classification of droplets in ddPCR images under variable exposure, contrast, and uneven illumination conditions without the need for image pre-processing and normalization processes.