Item
MartĂ BonmatĂ, Joan
MĂ©ridaudeau, Fabrice |
|
Universitat de Girona. Departament d’Arquitectura i Tecnologia de Computadors | |
Massich i Vall, Joan | |
2013 December 4 | |
This thesis analyses the current strategies to segment breast lesions in Ultra-Sound (US) data and proposes a fully automatic methodology for generating accurate segmentations of breast lesions in US data with low false positive rates. The proposed approach targets the segmentation as a minimization procedure for a multi-label probabilistic framework that takes advantage of min-cut/max- flow Graph-Cut (GC) minimization for inferring the appropriate label from a set of tissue labels for all the pixels within the target image. The image is divided into contiguous regions so that all the pixels belonging to a particular region would share the same label by the end of the process. From a training image dataset stochastic models are built in order to infer a label for each region of the image. The main advantage of the proposed framework is that it splits the problem of segmenting the tissues present in US the images into subtasks that can be taken care of individually En aquest treball, es proposa un sistema automĂ tic per generar delineacions acurades de lesions de mama en imatges d’ultrasò. El sistema proposat planteja el problema de trobar la delineaciĂł corresponent a la minimitzaciĂł d’un sistema probabilĂstic multiclasse mitjançant el tall de mĂnim cost del graf que representa la imatge. El sistema representa la imatge com un conjunt de regions i infereix una classe per cada una d’aquestes regions a partir d’uns models estadĂstics obtinguts d’unes imatges d’entrenament. El principal avantatge del sistema Ă©s que divideix la tasca en subtasques mĂ©s fĂ cils d’adreçar i desprĂ©s soluciona el problema de forma global |
|
application/pdf | |
http://hdl.handle.net/10803/128329 | |
eng | |
Universitat de Girona | |
Patologia. Medicina clĂnica. Oncologia
Enginyeria. Tecnologia |
|
Deformable object segmentation in ultra-sound images | |
doctoralThesis | |
TDX |