The Bhas42 cell transformation assay (Bhas42 CTA) is the only internationally certified assay that can evaluate carcinogenic promotion activity. However, the Bhas42 CTA requires the determination of the focus formed by cell transformation under a microscope, which requires sufficient proficiency and long hours of work by the experimenter. Therefore, this laboratory has been working on the creation of a convolutional neural network (CNN) model for the automatic focus determination of Bhas42 CTA. In a previous report, a CNN model (TPA model) that learned the focus formed by 12-O-Tetradecanoylphorbol-13-acetate (TPA), a positive control compound, was able to determine the focus with high accuracy, but its ability to determine the focus was reduced for other test substances. Therefore, in this study, we aimed to improve the generality of the CNN model by creating a TPA/LCA model with focus images of lithocholic acid (LCA) added to the training data. We also examined the effect of higher resolution image data sets and histogram matching on the judgment performance.

　When positive and negative judgments were made on the evaluation dataset using the TPA/LCA model, the judgment performance of the LCA image (percentage correct and AUC value) was significantly improved while maintaining the judgment performance of the TPA image. To confirm the effect of the resolution of the focus image on the judgment performance of the CNN model, we acquired focus images with low-resolution (1600 x 1200 px) and high-resolution (2880 x 2048 px) microscopes and subjected them to training and judgment. The results showed that when the resolution used for learning and judgment was different, the judgment performance was significantly reduced. To solve this problem, we introduced "histogram matching," which unifies the image data set with a histogram of pixel values of a reference image. By introducing histogram matching, we were able to create a CNN model that shows a correct response rate and AUC value of 0.80 or higher regardless of the image resolution.