Comparison of preprocessing methods for heart chambers segmentation from 2D CT images using 3D U-Net

Ovaj rad analizira metode predobrade pri uporabi za segmentaciju medicinskih slika srca dobivenih CT uređajem. Koristi metode promjene kontrasta izjednačavanje histograma, CLAHE, sigmoid correction i contrast stretching te metodu uklanjanja šuma median filter prije treniranja modela pomoću medicinskih slika pretvorenih iz 3D prostora (.nii) u 2D prostor (.jpg) kako bi model bio što točniji. Metoda strojnog učenja je U-net konvolucijska neuronska mreža koja ima veliku primjenu u segmentiranju medicinskih slika. Uz sve navedeno, rad pruža teorijsku osnovu svih korištenih metoda i tehnika te pojašnjava medicinsku pozadinu i uporabu slika. Na kraju rada provodi se analiza utjecaja spomenutih metoda predobrade te komentiraju rezultati.This paper analizes preprocessing methods used in case of medical image segmentation of human heart taken by CT device. It uses contrast enhancement methods izjednačavanje histograma, CLAHE, sigmoid correction and contrast stretching, and uses noise removing median filter before training model using medical images converted from 3D space to 2D space so that model would have higher accuracy. Machine learning method used is U-net convolutional neural network which has big applications in medical image segmentation. With all listed, this paper provides theoretical background of all used methods and tehniques while explaining medical background and image usage. At the end effect of methods is analized and results are commented