Extraction and evaluation of cell nuclei images in label-free phase contrast microscopy enabled by machine learning using a data analysis platform Usiigaci

Abstract

Recently, machine learning has been applied as a powerful tool for determining cell functions and states from cell shape features in microscopic images. However, it has not been fully investigated how changes in the microscopic observation environments and differences in cell types affect the accuracy of cell detection by machine learning. In addition to cell shapes, changes in intracellular structure are also considered important information in determining the cell functions. In particular, the shape of the cell nucleus is expected to be an indicator for disease diagnosis. In this study, we established a method for detecting the cell nuclear regions in phase contrast (PC) microscope images of unstained cells using Usiigaci (Tsai et al., 2019), an image machine learning platform consisting of Mask R-CNN. From a practical point of view, we investigated how the effects of out-of-focus state of the microscope and the cell type differences on the accuracy of nuclear region detection in PC microscope images. It was found that the accuracy of the nucleus segmentation was significantly influenced by the contrast (brightness gradient) of the captured images, and that an image taken at least within the depth of focus of the observation system is necessary for the accurate nucleus segmentation. Furthermore, the nucleus segmentation is easily affected by differences in cell morphology and intracellular structures such as the cytoskeletons, and the accuracy of the nucleus segmentation is less than half for vascular smooth muscle cells compared to NIH3T3 fibroblast cell lines. Therefore, compared to conventional segmentation for cell outlines, the nucleus segmentation using label-free PC microscope images is extremely difficult, and these are important issues that cannot be ignored in the practical application of machine learning of cell microscopic images.

Keywords

• cell biomechanics
• mechanobiology
• cell nucleus
• image analysis
• convolutional neural networks