Whole heart segmentation from MRI images using U-Net architecture

U ovom diplomskom radu, predstavljen je sustav za segmentaciju 2D MRI (engl. Magnetic Resonance Image) slika srčanih komora. Za segmentaciju korištena je U-net konvolucijska neuronska mreža. Kroz rad dana je klinička pozadina o strukturi srca i fizikalna pozadina rada uređaja magnetske rezonancije. Opisani su formati zapisa slika magnetske rezonancije i način rada konvolucijskih neuronskih mreža. Sustav primjenjuje posebnu metodu obrade slike temeljena na Bayesovom teoremu. Evaluacija rezultata uspoređuje rezultate segmentacije sa i bez predobrade podataka. Provedena evaluacija postiže bolje rezultate sa predobradom podataka, ali postoje slučajevi kada zbog prejake predobrade podataka, rezultati nisu uvijek zadovoljavajući.In this thesis, a system for segmentation of 2D MRI images of heart chambers is presented. The U-net convolutional neural network was used for the segmentation process. The clinical background on the structure of the heart and the physical background of the operation of MRI device were given throughout the work. Magnetic resonance image recording formats and convolutional neural networks are described. The system applies a special image processing method based on Bayes' theorem. Results evaluation compares segmentation results with and without data preprocessing. The conducted evaluation achieves better results with data pre-processing, but there are cases when due to too strong data pre-processing, the results are not always satisfactory