A novel method based on a Mask R-CNN model for processing dPCR images

A digital polymerase chain reaction (dPCR) using fluorescence images for collecting quantitative information needs efficient software tools to automate the image analysis process. However, due to the broad range of fluorescence image characteristics, such as the uneven fluorescence intensity distribution, irregular structures of microarrays, and diverse and unpredictable morphologies of microchambers, existing tools fail to extract signals from these images properly, thus posing challenges for the improvement of detection accuracy. In this paper, a deep learning method based on the Mask R-CNN model was used for image processing to achieve more accurate quantification of nucleic acids in both microarray and droplet dPCR. This Mask R-CNN based method uses massive dPCR fluorescence image data to train a model that has the ability to recognize target signals in dPCR images precisely and automatically, regardless of the non-uniform luminosity or spot impurities appearing in dPCR images. When the Mask R-CNN model is used to process images with non-uniform luminosity, the true positive rate of this model can reach 97.56%. By contrast, the true positive rate of threshold segmentation is only 68.29%. As for dealing with images with spot impurities, which caused a 7.25% fault of the positive points in the threshold segmentation method, the error rate decreased to zero using the Mask R-CNN model. In addition, the labeling of annotated pictures is time-consuming. Therefore, an iteration method was used in the annotation stage to reduce time spent on labeling. With proper modifications, this method has great potential to serve as an alternative to achieve a highly efficient fluorescence image process for dPCR or other digital assays.