Rapid and sensitive mycoplasma detection system using image-based deep learning

A major concern in the clinical application of cell therapy is the manufacturing cost of cell products, which mainly depends on quality control. The mycoplasma test, an important biological test in cell therapy, takes several weeks to detect a microorganism and is extremely expensive. Furthermore, the manual detection of mycoplasma from images requires high-level expertise. We hypothesized that a mycoplasma identification program using a convolutional neural network could reduce the test time and improve sensitivity. To this end, we developed a program comprising three parts (mycoplasma detection, prediction, and cell counting) that allows users to evaluate the sample and verify infected/non-infected cells identified by the program. In experiments conducted, stained DNA images of positive and negative control using mycoplasma-infected and non-infected Vero cells, respectively, were used as training data, and the program results were compared with those of conventional methods, such as manual counting based on visual observation. The minimum detectable mycoplasma contaminations for manual counting and the proposed program were 10 and 5 CFU (colony-forming unit), respectively, and the test time for manual counting was 20 times that for the proposed program. These results suggest that the proposed system can realize a low-cost and streamlined manufacturing process for cellular products in cell-based research and clinical applications.