Návrh a realizace softwaru pro automatizovanou analýzu mozkových aktivací z funkční magnetické rezonance pomocí umělé inteligence

Abstract

In this work, artificial intelligence, specifically convolutional neural networks, are used to automatically threshold activations of functional magnetic resonance images. Functional magnetic resonance (fMRI) is a modern imaging method for mapping the gray matter of the cerebral cortex. During an fMRI acquisition, a volunteer or a patient, performs a specific task, e.g. movomenet of of the left upper limb, during which the changing oxygenation in different parts of the brain tissue is measured by the MRI scanner. However, these activations require subsequent, time-consuming, processing. This processing is furthermore burdened by subjective factors. In this work, an entirely new methology for so-called thresholding of fMRI images is proposed to reduce the time-consuming nature of overall processing and to remove the subjective processing factor. The U-Net architecture was chosen for automatic thresholding and the nnU-Net framework was used for implementation. The resulting trained neural network averages a DICE coefficient of 0,777. Furthermore, the Hausdorff distance, accuracy, sensitivity and relative volume error (RVE) metrics were used for quantification. The values mentioned were 4,46 (mm); 0,755 (-); 0,684 (-) and 25,7 (%), respectively. Objectively, the architecture can be used to automatically threshold different fMRI data.