Random Walker For Prostate Segmentation With Automated Seed Generation In Mri

Prostate cancer is the second driving reason for malignant growth in deaths among men,which pushes scientists to search for promising methods to decrease the loss of human's life.Medical imaging contributes a lot in the therapeutic discipline due to the fact of its wide use in diseases analysis and treatment of patients.Ultrasound,computed tomography,magnetic resonance imaging(MRI),and positron emission tomography are typical modalities of medical imaging and most of them commonly suffer from noise and sampling artifacts.The interpretation of results of these modalities relies primarily on the perception of the radiologists.Image segmentation of the prostate in(MRI)aims to figure out which area in an MRI image contains the prostate.Finding prostate boundaries in images is very useful in a large number of applications and is of much significance when directing prostate investigation for volume calculation,malignant growth stages understanding,treatments and patients check out,etc.However,segmentation of the prostate in MRI images is challenging due to the inhomogeneous intensity variations appearing during digital image acquisition and its variation in shape.Because of those reasons,outlining the prostate shape is an incredibly long and challenging task,and therefore frequently an obstacle in many research projects.In this thesis,I investigated prostate segmentation based on a popular graph-based random walker(RW)segmentation algorithm that incorporates the intensity features in prostate images and its shape variability to overcome challenges.The ability of this method relies on its robustness with respect to incomplete contours and its efficient optimization.The RW algorithm was developed for interactive segmentation allow the user to pre-segment small region called seeds in the foreground and background of the prostate region.From the seeds,the algorithm computation would then carry out the entire segmentation based on graph theories.The work had two major stages;In the first part of my methodology,I selected slices with visible prostate contour and generated their ground truth to be used for Atlas generation.During this process,the ground-truth images were sampled to the same pixel size and same image size.Then,all ground-truth images were stacked together into a single image space.After the stacking process,a probability map was generated in that image space and two probability levels were chosen to extract the inner seed(foreground seed)and the outer seed(background seed).Those seeds generated in atlas were considered as automated by initialized seeds for RW method.The second part of my algorithm was the use of normal RW process by generating the seeds manually.The manual seeds were well placed inside the target region and the RW algorithm would segment the contours accurately.Through a sampling stage of this method,unlabeled pixels were generated within the image domain and each pixel was taken into consideration as a node and was accredited to a label.The weight of the edges between the seeds was then set to provide information about the existence or non-existence of the prostate contour between two seeds and the similarities between the edges.Both seed initialization methods provided good results when compared their segmentations with the ground truth images.The segmentation obtained averaged results of 0.81 on Dice similarity coefficient and 0.68 on Jaccard coefficient in automatic seed initialization,while in manual seed initialization the results increased to 0.88 on Dice similarity coefficient and 0.75 on Jaccard coefficient,which indicates successful segmentation.