Lung Segmentation Using Multi-atlas Registration and Graph Cuts

Segmentering af lunger vha. multi-atlas registrering og graph cuts

Sandberg, Thomas Holm

4. term

Biomedical Engineering and Informatics, Master

2014

2014-06-03

70

CT-billeder af lungerne bruges til at diagnosticere og følge emfysem hos patienter med KOL, men automatiseret analyse kræver en præcis afgrænsning af lungerne. Denne afhandling præsenterer en fuldautomatisk segmenteringsmetode, der kombinerer multi‑atlas registrering (for at indfange global forminformation) med et graph‑cut framework, hvor voxel‑intensitet og lokale naboforhold indgår i en samlet optimering. Et sandsynlighedskort for lungevæv dannes ud fra registrerede atlasser og bruges som rumlig prior sammen med en intensitets- og nabomodel til at bygge en graf, hvis optimale opdeling beregnes ved graph‑cut som en maksimum a posteriori/energi‑minimering. Metoden blev evalueret på 19 bryst‑CT‑scanninger i et leave‑one‑out‑setup med Dice similarity coefficient (DSC) og sammenlignet med en multi‑atlas segmenteringsmetode, samt afprøvet på et HRCT‑datasæt med 10 raske og 10 patienter med emfysem ved visuel og kvantitativ vurdering, hvor manuelle referenceindtegninger forelå. Graph‑cut metoden opnåede en højere DSC end multi‑atlas i krydsvalideringen (0,9877 vs. 0,9858), segmenterede succesfuldt lungerne på HRCT‑datasættet og viste bedre ydeevne end multi‑atlas, hvor manuelle sammenligninger var mulige. Resultaterne er på niveau med andre lungesegmenteringsstudier, men yderligere data fra patienter med emfysem og andre lungesygdomme er nødvendige for direkte sammenligning og yderligere vurdering. Metoden giver en nøjagtig, objektiv og arbejdsbesparende segmentering, der kan danne grundlag for efterfølgende computerbaserede analyser af lungerne.

Chest CT is widely used to diagnose and monitor emphysema in COPD, but reliable automated analysis depends on accurate lung delineation. This thesis presents a fully automatic segmentation method that combines multi‑atlas registration (to capture global anatomical shape) with a graph‑cut framework that integrates voxel intensity and local neighborhood information into a single optimization. A spatial probability map derived from registered atlases serves as a prior alongside intensity and neighborhood models to build a graph, and the optimal segmentation is obtained via graph‑cut as a maximum a posteriori/energy minimization. The method was evaluated on 19 chest CT scans in a leave‑one‑out setup using the Dice similarity coefficient (DSC) and compared with a multi‑atlas segmentation approach, and further tested on an HRCT dataset of 10 healthy subjects and 10 patients with emphysema by visual and quantitative assessment where manual annotations were available. The graph‑cut method achieved a higher DSC than the multi‑atlas approach in cross‑validation (0.9877 vs. 0.9858), successfully segmented the lungs on the HRCT dataset, and outperformed the multi‑atlas method where manual comparisons were possible. Performance is comparable to reports in the literature, though more data from emphysema and other lung diseases are needed for direct comparison and further evaluation. The method provides accurate, objective, and labor‑saving segmentation that can support downstream computational analyses of the lungs.

[This summary has been generated with the help of AI directly from the project (PDF)]

segmentation ; multi-atlas ; atlas ; registration ; graph cuts ; ct ; imaging ; image ; image registration ; image processing ; lungs ; copd ; emphysema

Download
View record in AAU Student Projects