Segmentacija CT slik celotne hrbtenice in uporaba v praksi : diplomsko delo

Pavošev, Michel (author), Vuga, Rok (author), Žibert, Janez (mentor), Žager Marciuš, Valerija (reviewer)

Uvod: Segmentacija je temeljni postopek pri analizi medicinskih slik, ki vključuje prepoznavanje in natančno anotacijo področij interesa. Ločimo ročno segmentacijo, polavtomatsko in avtomatsko. MONAI Label je inteligentno orodje zasnovano na podlagi globokih nevronskih mrež, ki uporabniku omogoča označevanje medicinskih slik, pri čemer so označevalne aplikacije dostopne preko strežnika. Namen: Cilj diplomskega dela je primerjati hitrost in učinkovitost segmentacijskih metod na CT slikah celotne hrbtenice, kar bi omogočilo natančnejšo in lažjo vizualizacijo vretenc za uporabo v klinični praksi. Pri tem bomo uporabili program 3D Slicer ter orodja MONAI Label, Auto3Dseg in segmentacijo s povečevanjem območja. Metode dela: Opravili smo pregled literature, kjer smo iskali vire, na področju segmentacije in njenih metod, pretežno v angleškem jeziku in uporabo tujih podatkovnih baz. Temu je sledila študija primera, pri čemer smo uporabili javno dostopno arhivsko zbirko TCIA za prenos CT slik. Iz celotne zbirke smo izbrali in segmentirali prvih šestnajst preiskav ABD_LYMPH, ki zajemajo področje mediastinuma in abdomna ter tako zagotovili, da je bil na sliki zajet celoten del ledvene in torakalne hrbtenice ter del vratne hrbtenice. Izvajalec ena in izvajalec dva smo, CT posnetke segmentirali na enak način z uporabo polavtomatske in avtomatske segmentacije z orodji Auto3Dseg. Po opravljeni segmentaciji smo uporabili razširitveni vtičnik Slicer RT za izračun koeficienta ujemanja ter primerjali vse rezultate. Naredili smo še statistično analizo rezultatov s t-testom odvisnih vzorcev in ANOVO. Rezultati: Podatke polavtomatske in avtomatske segmentacije modelov Vertebrae TS2 in Whole body segmentation TS2 smo razdelili na porabljen čas in izračunan Dice koeficienta, v drugem delu pa smo rezultate med seboj primerjali, jih grafično prikazali in ugotovili ali obstajajo statistično značilne razlike. Razprava in zaključek: Iz rezultatov smo ugotovili, da je najhitrejši ter natančnejši segmentacijski model Vertebrae TS2. Avtomatska metoda segmentacije je najhitrejši način označevanja slik, v primerjavi s polavtomatsko metodo. Implementacija v praksi bi lahko prinesla spremembe tako na področju izobraževanja, diagnostike in terapije, vendar je potrebno razviti učne strategije.

• Language: Slovenian
• Keywords: diplomska dela, radiološka tehnologija, TCIA, MONAI Label, metode segmentacije, 3D Slicer, Auto3DSeg
• Work type: Bachelor thesis/paper
• Typology: 2.11 - Undergraduate Thesis
• Organization: ZF - Faculty of Health Sciences
• Place of publishing: Ljubljana
• Publisher: [M. Pavošev : R. Vuga]
• Year: 2026
• Number of pages: IX, 41 str.
• PID: 20.500.12556/RUL-186564
• UDC: 616-07
• COBISS.SI-ID: 289894659
• Publication date in RUL: 03.09.2026

Secondary language

• Language: English
• Title: Whole-spine CT image segmentation and its implementation in practise : diploma work

Abstract

Introduction: Segmentation is a fundamental process in medical image analysis involving the identification and precise annotations of regions of interest. Manual, semi-automatic, and automatic segmentation are distinguished. MONAI Label is an intelligent tool based on deep neural networks that enables the user to annotate medical images, with labeling applications accessible via a server. Purpose: The aim of this diplomas thesis is to compare the speed and efficiency of different segmentation methods on CT images of the entire spine, enabling more accurate and easier visualization of vertebrae for clinical practice. For this purpose, we used the 3D Slicer software along with the MONAI Label tools, Auto3Dseg, and region-growing segmentation. Methods: We conducted a literature review, searching primarily English-language sources in international databases on segmentation and its methods. This was followed by a case study in which we used the publicly available TCIA archive to download the CT images. We selected and segmented the first sixteen ABD_LYMPH examinations, covering the mediastinum and abdomen, ensuring that the images included the entire lumbar and thoracic spine as well as part of the cervical spine. Operator one and operator two segmented the CT scans in the same way using semi-automatic and automatic segmentation with the Auto3Dseg tools. We later used the SlicerRT extension to calculate the Dice similarity coefficient and compared all results. We also performed a statistical analysis of the results using a paired-samples t-test and ANOVA. Results: The data from the semi-automatic and automatic segmentation models (Vertebrae TS2 and Whole Body Segmentation TS2) were divided by the processing time and calculated Dice coefficient. In the second part, the results were compared, presented graphically and we later determined whether statistically significant differences existed. Discussion and conclusion: From our results, we concluded that the Vertebrae TS2 segmentation model is the fastest and most accurate. The automatic segmentation method labels images more quickly than the semi-automatic method. Its implementation in practice could bring changes to education, diagnostics, and therapy; however, it is necessary to develop appropriate learning strategies.

• Keywords: diploma theses, radiologic technology, TCIA, MONAI Label, segmentation methods, 3D Slicer, Auto3Dseg