Research On Digital PCR Dense Droplets Detection Based On Hybrid Deep Learning Model

Digital Polymerase Chain Reaction(dPCR)is a highly sensitive technique for the absolute quantitative detection of nucleic acid.Its principle is to disperse the DNA template into a large number of droplet microunits for specific DNA template amplification reaction,and count the number of fluorescence droplets,so as to determine the copy number of the initial DNA template.At present,the main steps of the mature fluorescence PCR detection systems including scanning droplets using laser beam,collecting and focusing to the photoelectric detector,and analyzing the PCR fluorescence signal of the droplet.However,these methods are easily influenced by the subjective factors,and have poor detection accuracy and low efficiency,which can not meet the social needs of high efficiency and high yield.In this study,to further improve the detection performance of the PCR droplet method,we focus on the extraction of Region of Interest(RoI)from the PCR droplet image,extraction of deep discriminant features.The main contributions are as follows:(1)In view of the problems such as large number of invalid areas and too few data samples in the collected PCR droplet images,the images should be preprocessed before intensive droplet detection.The preprocessing steps included the use of image graying,binarization,linear detection and other techniques to extract RoI quickly and image enhancement to expand data.The preprocessing process avoided the repetitive detection of droplet and improved the reliability of training.(2)A co-occurrence flow enhanced bidirectional pyramid convolution dense droplet detection method was constructed to solve the problems of small targets and relatively large changes in physical structure and geometric appearance of PCR droplets.This study introduced a variety of data enhancement methods to enhance the target droplet in the droplet image to make the target information clearer.First,to strengthen the correlation between the pyramid network and layers,a two-way pyramid convolutional network with temporal and spatial branches was designed to capture the droplet deep discrimination features.Secondly,the low-level physical appearance,high-order semantic and global contextual information of droplet were modeled,and the slice co-occurrence attention network was established to further enhance the representation of target droplet by information flow.Finally,the multi-level information such as low level,middle level and high level was cross-aggregated to detect the droplet accurately.In addition,to improve the robustness of the network in the feature extraction stage,a combined loss function was designed to adjust and optimize the network to further improve the droplet detection accuracy.(3)In order to solve the problems of fuzzy boundary division between adjacent droplets in PCR droplet image,dense target,and few samples,a multi-model fusion dense droplet detection algorithm composed of dynamic aggregation network and multi-scale feature alignment network was chosen.Firstly,a variety of feature extractors were used to model the low-level and high-level information of droplet.Secondly,the obtained features were fused to form complementary features among different network features to achieve accurate droplet detection.Finally,in order to demonstrate the effectiveness of the proposed method,the PCR droplet dataset was evaluated and validated to ensure the accuracy and detection efficiency.This study was a fast and robust PCR droplet detection technology,which would play a positive role in the clinical application and promotion of photographic digital PCR technology.