Evaluation and analysis of data driven in expectation maximization segmentation through various initialization techniques in medical images
Autor: | R. Varatharajan, A. Karthikeyan, V. Prabhu, P. G. Kuppusamy |
---|---|
Rok vydání: | 2018 |
Předmět: |
Mean squared error
Computer Networks and Communications Computer science Initialization 02 engineering and technology 030218 nuclear medicine & medical imaging 03 medical and health sciences 0302 clinical medicine Histogram Expectation–maximization algorithm 0202 electrical engineering electronic engineering information engineering Media Technology medicine Segmentation Cluster analysis medicine.diagnostic_test business.industry 020207 software engineering Magnetic resonance imaging Pattern recognition Image segmentation medicine.disease Hardware and Architecture Outlier Artificial intelligence business Software Glioblastoma |
Zdroj: | Multimedia Tools and Applications. 77:10375-10390 |
ISSN: | 1573-7721 1380-7501 |
Popis: | The operation of partitioning an image into a collection of connected sets of pixels is known as image segmentation. Expectation Maximization (EM) segmentation requires appropriate initialization of tissue class mean and variance, as it is get stuck on particular possible area of intensity of the probability nature. Decisive initialization is a necessary exploratory process for the forthcoming convergence of the algorithm to best regional maximum of possibility task. Indiscriminate initialization is not dossier directed, deep from optimal, outcomes are not reproducible, do not take advantage of deep rooted patterns in the data or may be loaded on outliers. This paper evaluates the performance of EM segmentation with random initialization, histogram guided initialization and initialization with k-means with respect to computational complexity and root mean squared error of tissue class mean and variance, updated by EM, with manually estimated tissue class mean and variance as ground truth, on paramount plane, T1 contrast and Magnetic resonance (MR) images of Glioblastoma Multiframe (GBM) Edema complex. The random initialization and histogram guided initialization was experimented for k-means, from the clustered output of which; initial tissue class mean and variance for EM are derived. RMS error remains the same for EM initialized and histogram guided K-means. EM initialized with K-means which has histogram guided initialization converges fast than the random initialization K-means, but the computational time is more for the former initialization than the latter. The experimental evaluation of EM initialization schemes and Fuzzy Cluster Means (FCM) were performed in MATLAB. The efficacy of Fuzzy Cluster Means clustering was analyzed qualitatively on tumour-edema complex. FCM could identify only 3 classes including background in the MR specimens. FCM consider edema and certain parts of WM as a single tissue class. Similarly, FCM clubs GM, CSF and necrotic focus into tissue class and produced empty clusters. |
Databáze: | OpenAIRE |
Externí odkaz: |