Segmentation of Fungal Cells from Multiplane Brightfield Microscopy Images
The analysis of digitally segmented cell populations from microscopy images has a critical part in cell quantification and phenotyping. Brightfield microscopy gives an accessible method for quick visualization of cells and is becoming more frequently used for cell image segmentation. Using multiple brightfield microscopy images from different focal planes has been proven to benefit cell segmentation even further. The booming development of Deep Learning algorithms for image segmentation relying on Convolutional Neural Networks facilitated automated cell segmentation at a high throughput speed and with improved sensitivity. In this thesis, we approached the two-dimensional segmentation of fungal cells - a widely underrepresented cell morphology distinguished by elongated and branched cell shapes - using three-dimensional brightfield image stacks. We modified the original U-Net implementation and trained a cell segmentation model in a fully supervised way with the custom-made ground-truth segmentation masks obtained from three-dimensional stacks of fluorescence microscopy images. We additionally examined through a series of experiments with full-stack, single-plane, and two-plane segmentation how the cell segmentation performance is compromised depending on the number of input image planes provided to the model for training. Based on the obtained results, we observed that the multiplane approach to fungal cell segmentation does not always lead to improved performance over the single-plane segmentation. However, when selecting two complementary focal planes, we were able to match the performance level of the full-stack segmentation, achieving a 2% relative difference in the score for the majority of used evaluation metrics.
Graduation Thesis language
Graduation Thesis type
Master - Computer Science