HE_I_L: Health Sciences I. Lectures
Gyula Ujlaki, University of Debrecen, Faculty of Medicine, Department of Medical Chemistry, Debrecen
Introduction: With the fast-developing high throughput microscopy techniques and the growing size of digital image databases, the demand for automatic image analysis techniques is increasing. Regarding the automatized image analysis techniques, we generally distinguish between two different approaches. Conventional image analysis pipelines comprise multiple steps, each step requiring method customization and parameter adjustments. In contrast to this, machine learning methods do not require further intervention after the training sessions. Our pilot experiments aimed at developing new microscopy image segmentation models and analysis methods, using convolutional neural networks (CNN) and conventional pipelines combined, in in vitro experiments with cancer cells.
Aims: To develop high throughput and fully automatized image analysis for high throughput microscopy.
Methods: The applied techniques were immunocytochemistry, high throughput microscopy, machine learning and image analysis.
Results: The training datasets were consisting of both brightfield and fluorescently labeled, high throughput microscopy images of various cancer cell line types. Although the trained neural network is capable segmenting the cancer cells on brightfield images, regardless of cell origin, training the neural network for fluorescently labeled image segmentation was only possible if the training set only contained cells with similar morphology. It was not possible with the currently available training dataset and parameter adjustments, to create a multiclass segmentation model to process both brightfield and fluorescently labeled images. Taking the advantage of conventional image analysis pipeline tools, data extraction of the segmented objects can also be done.
Conclusion: In this project, we created image segmentation models and developed automatic analysis methods specifically for cancer cell segmentation and image data extraction from brightfield and fluorescently labeled images.
Funding: The work is supported by the EFOP-3.6.1-16-2016-00022 project. The project is co-financed by the European Union and the European Social Fund.
Debrecen University Faculty of Medicine, Doctoral School of Molecular Medicine