Modeling of data from 3D medical images

U ovom radu opisano je nekoliko metoda segmentacije volumena te njihova primjena na konkretan zadatak izrade trodimenzionalnog modela odabranog objekta iz medicinskih CT snimaka dijela ljudskog tijela. Opisan je proces nastanka medicinskih slika u DICOM formatu koje služe kao ulazni podaci, te njihovo oblikovanje za potrebe daljnje obrade. Iz niza 2D slika koje predstavljaju presjeke kroz tijelo odabrano je područje interesa koje sadrži objekt koji se želi segmentirati, u ovom slučaju nekoliko kralježaka. Iz tog područja formirana je trodimenzionalna diskretna rešetka poštujući realne dimenzije koje su zapisane u DICOM metapodacima. Implementirane su tri metode formiranja 3D modela: amplitudna segmentacija koja se temelji na pronalasku područja iznad zadanog praga, te dvije metode koje se temelje na pronalasku graničnih ploha, proizašle iz metoda za detekciju bridova na 2D slici --- segmentacija pomoću Cannyjevog algoritma te segmentacija temeljena na lokalnoj faznog koherenciji. Kako za objekt segmentacije u ovom radu nema fizičkog referentnog modela, za svaku metodu formiran je veći broj modela varijacijama u parametrima koje algoritmi koriste, te je provedeno usrednjavanje s ciljem dobivanja minimalne ukupne pogreške u procjeni, te tako i što boljeg modela.In this thesis several methods of volume segmentation have been described, and also their application to the specific problem of constructing a three-dimensional model of a selected object from medical CT images of the human body. The thesis also describes the process of formation of medical images in DICOM format which are used as input data, and their shaping for further processing. A region of interest which includes the object of segmentation is selected from the series of 2D images that represent axial cross-sections through the body. The object of segmentation are several vertebrae. A 3D discrete grid is formed from the region of interest, considering the real-world dimensions which are stored in DICOM metadata. Three methods for 3D modeling are implemented: amplitude thresholding based on finding the regions with values above a threshold, and two methods based on finding a boundary surface (developed from methods for edge detection on a 2D image) --- Canny algorithm segmentation, and segmentation based on local phase coherence. As there is no physical referent model for selected object of segmentation, for every method a number of models has been formed by variating the algorithm parameters. Using those models, averaging has been conducted in order to find an estimation of model with minimal dispersion