engleski

Application of convolutional neural network for protein counting and localization photographed by super-resolution microscopy

In recent years novel microscopy techniques have led to the development of super-resolution imaging with increased resolution compared to conventional microscope techniques. A branch of super-resolution microscopy, based on single-molecule localization that includes techniques such as Photoactivated localization microscopy (PALM) and Stochastic Optical Reconstruction Microscopy (STORM), enables researchers to study cellular processes with increased resolution. Single-molecule localization microscopy relies on collecting a set of images where only a subset of optically resolvable fluorophores is emitting light. The analysis of these images allows for reconstruction of high-quality super-resolution images. This graduation thesis proposes a further step in super-resolution image analysis, based on the analysis of individual protein localization. It proposes to use the multiple counts produced by a single protein to predict its position. Position is predicted beyond overcounting artifact through a neural network, in order to predict the protein number and position. Counting proteins from super-resolution image is a problem because of fluorophore overcounting. Unlike previous papers trying to solve this problem, this graduation thesis does not use any modelling and relies only on neural network's ability to learn from data. This approach provides researchers with a fast performing algorithm that requires minimal knowledge about neural networks and image analysis. The main object of this graduation thesis is to create and measure how well the state of the art image processing neural network can localize and count proteins.

super-resolution microscopy, CNN, machine learning, protein counting and localization

nije evidentirano