The clinical effectiveness of intraoperative flow cytometry (iFC) for glioma has been studied by several researchers. However, the studies of its analytic validity are yet few. Particularly, the repeatability, which is a very basic characteristic of analytic validity, cannot be evaluated in principle because the specimen becomes a cell suspension by cell dispersion. Therefore, in this study, a stochastic measurement model was developed and a surrogate index to indirectly evaluate the repeatability was defined. In addition, the cell dispersion function of Celltac PEAK, a clinical flow cytometer dedicated to iFC, was evaluated by this method. As a result, the repeatability of the cell dispersion function was interval-estimated as [0.2%, 1.7%] with standard deviation.

Artifi cial intelligence has gained massive attention in research, thanks to extensive development of algorithms and continuing growth of computing power. Particularly, deep learning represented by a Convolutional Neural Network has enabled high-accuracy image recognition compared with conventional machine learning methods. To classify images using deep learning, the most popular method is pre-learning, also called "supervised learning." This method uses labeled training datasets before classifi cation of test images. Nowadays, many researchers are implementing machine learning in automatic analysis of medical images. In life sciences, however, preparing such labeled training datasets requires huge effort in time and cost. We describe a strategy to obtain training image datasets of cells using highthroughput imaging fl ow cytometry. We found that the accuracy of cell type classifi cation was signifi cantly higher using machine learning with image data sets acquired by imaging fl ow cytometer than that of human capability. Here we show that it is also possible to apply this approach in unsupervised learning. Artifi cial intelligence has gained massive attention in research, thanks to extensive development of algorithms and continuing growth of computing power. Particularly, deep learning represented by a Convolutional Neural Network has enabled high-accuracy image recognition compared with conventional machine learning methods. To classify images using deep learning, the most popular method is pre-learning, also called "supervised learning." This method uses labeled training datasets before classifi cation of test images. Nowadays, many researchers are implementing machine learning in automatic analysis of medical images. In life sciences, however, preparing such labeled training datasets requires huge effort in time and cost. We describe a strategy to obtain training image datasets of cells using highthroughput imaging fl ow cytometry. We found that the accuracy of cell type classifi cation was signifi cantly higher using machine learning with image data sets acquired by imaging fl ow cytometer than that of human capability. Here we show that it is also possible to apply this approach in unsupervised learning.