Level-set appearance modeling for segmentation of anatomical structures in 3-D medical images
Author
Stephansen, Ulrik
Term
4. term
Publication year
2012
Submitted on
2012-06-01
Pages
102
Abstract
Medicinsk billeddannelse bruges stadig mere, og behovet for automatisk segmentering af relevante strukturer vokser. I dette arbejde implementerer vi en modelbaseret metode: et Active Appearance Model (AAM), som lærer typiske former og udseender ved hjælp af hovedkomponentanalyse (PCA). Formen repræsenteres som et level-set, og modellen tilpasses iterativt til 3D-billeder. Algoritmen bruger forhåndsviden til at forudsige, hvordan model- og poseparametre (størrelse, placering og orientering) skal justeres for at opnå en bedre tilpasning til et nyt billede. Metoden afprøves på 42 MR-billeder af prostata og 27 CT-billeder af L4-hvirvlen i en leave-one-out-krydsvalidering (træning på alle undtagen ét billede, gentaget for alle). De automatiske segmenteringer sammenlignes med manuelle reference-segmenteringer. Den opnåede mediane Dice-overlapscore er 0,81 for begge strukturer (1,0 er perfekt overlap). Ydelsen er på niveau med tidligere metoder, men algoritmen fastlægger i nogle tilfælde prostatas størrelse forkert. Desuden er udseendemodellen ikke stor nok til fuldt at segmentere hvirvelprocesserne. Algoritmen er følsom over for den indledende placering af gennemsnitsmodellen i mål-billedet. Den præsenterede AAM-tilgang kan anvendes på forskellige billedmodaliteter og strukturer, hvis formen ikke varierer for meget.
Medical imaging is increasingly used, and there is a growing need to automatically outline structures of interest. This work implements a model-based method: an Active Appearance Model (AAM) that learns typical shape and appearance using principal component analysis (PCA). Shapes are represented with a level-set surface, and the model is fitted to 3D images through an iterative procedure. The algorithm uses prior knowledge to predict how to adjust model and pose parameters (size, position, and orientation) to better match a new image. We test the method on 42 prostate MRI scans and 27 CT scans of the L4 vertebra using leave-one-out cross-validation (training on all but one image, repeated for each case). The automatic segmentations are compared with manual references. A median Dice overlap score of 0.81 is achieved for both structures (1.0 indicates perfect overlap). Performance is comparable to previous methods, but the algorithm sometimes estimates the prostate size incorrectly. The appearance model is also not large enough to fully segment the vertebral processes. The method is sensitive to the initial placement of the average model in the target image. The presented AAM can be applied to different imaging modalities and structures if the shape does not vary too much.
[This abstract was generated with the help of AI]
Documents