Biomedical Image Processing / Medical Image Processing
Gelareh Valizadeh; Farshid Babapour Mofrad; Ahmad Shalbaf
Volume 14, Issue 4 , February 2021, , Pages 291-306
Abstract
Statistical Shape Modeling is widely used in many applications of cardiac images. Many efforts have been done to generate optimized Statistical Shape Models (SSMs). In this paper, we evaluated three different 3D endocardial models constructed using different alignment procedures. From 20 healthy CMR ...
Read More
Statistical Shape Modeling is widely used in many applications of cardiac images. Many efforts have been done to generate optimized Statistical Shape Models (SSMs). In this paper, we evaluated three different 3D endocardial models constructed using different alignment procedures. From 20 healthy CMR datasets, three different endocardial models are generated by varying the surface alignment methods means based on the Center of the Apex (CoA), the Center of Mass (CoM), and the Center of the Basal (CoB) of the endocardium. Then Principle Component Analysis (PCA) is applied to show the maximum variation of the SSMs. The constructed statistical models are compared by measuring the compactness, generalization ability, and specificity. Besides, the performance of each model in the 3D endocardial segmentation application using the Active Shape Model (ASM) technique is evaluated by the Hausdorff Distance (HD) criterion. The results indicate that the CoB-based model is less compact than the CoA-based model but more compact than the CoM-based model. Although for a constant number of modes the reconstruction error is approximately the same for all models, surface alignment based on CoB leads to generate a more specific model than the others. The resulted HDs show that the CoB alignment strategy produces the ASM which has the best performance in 3D endocardial segmentation among the other models. The computed results from the quantitative analysis demonstrate that varying alignment strategies affect the quality of the constructed SSM. It is obvious that the specificity and segmentation accuracy of the proposed CoB-based model outperforms the classical CoM-based approach.
Biomedical Image Processing / Medical Image Processing
Neda Behzadfar; Hamid Soltanian Zadeh
Volume 7, Issue 3 , June 2013, , Pages 219-236
Abstract
Segmentation of tumors in magnetic resonance images is an important task. However, it is quite time consuming and has low accuracy and reproducibility when performed manually. Automating the process is challenging, due to high diversity in appearance of tumor tissue in different patients and in many ...
Read More
Segmentation of tumors in magnetic resonance images is an important task. However, it is quite time consuming and has low accuracy and reproducibility when performed manually. Automating the process is challenging, due to high diversity in appearance of tumor tissue in different patients and in many cases, similarity between tumor and normal tissues. This paper presents semi-automatic approach for analysis of multi-parametric magnetic resonance images (MRI) to segment a highly malignant brain tumor called Glioblastoma multiform (GBM). MRI studies of 12 patients with GBM tumors are used. To show that the proposed method identifies Gd-enhanced tumor pixels from T1-post contrast images minimal user interactions. They are also used to illustrate that the segmentation results obtained by the proposed approach are close to those of an expert, by showing excellent correlations among them (R2=0.97). In order to evaluate the proposed method in practical applications, effects of treatment of GBM brain tumors using Bevacizumab are predicted. Bevacizumab is a recent therapy for stopping tumor growth and even shrinking tumor through inhibition of vascular development (angiogenesis). To this end, two image series of 12 patients before and after treatment and relative changes in the volumes of the Gd-enhanced regions in T1-post contrast images are used as measure of response. The proposed method applies signal decomposition with KNN classifier to minimize user interactions and increase reproducibility of the results. Then histogram analysis is applied to extract statistical features from Gd-enhanced regions of tumor and quantify its micro structural characteristics. Predictive models developed in this work have large regression coefficients (maximum R2=0.91) indicating their capability to predict response to therapy. The results obtained by the proposed approach are compared with those of previous work where excellent correlations are obtained.
Biomedical Image Processing / Medical Image Processing
Abbas Biniaz; Ataollah Abbasi; Mousa Shamsi
Volume 7, Issue 2 , June 2013, , Pages 175-186
Abstract
Segmentation divides an image to some subdivisions where which of ones has similar intensity gray levels. Among clustering methods fuzzy c-means (FCM) clustering has been frequently used for segmentation of medical images. However, this algorithm doesn’t incorporate spatial neighborhood information ...
Read More
Segmentation divides an image to some subdivisions where which of ones has similar intensity gray levels. Among clustering methods fuzzy c-means (FCM) clustering has been frequently used for segmentation of medical images. However, this algorithm doesn’t incorporate spatial neighborhood information in segmentation. This approach is very susceptible to nuisance factors. Therefore this paper proposes a Gaussian spatial FCM (gsFCM) to MR image segmentation. Proposed method has less sensitivity to noise specially in tissue boundaries, angles, and borders than spatial FCM (sFCM). Furthermore by the suggested algorithm a pixel which is a separate tissue from structurally point of view for example a tumor in primary stages of its appearance, has more chance to be a unique cluster. Applying quantitative assessments using Jaccard similarity index, Dice coefficient, and other validation functions on FCM,sFCM and gsFCM approaches show efficient performance of the proposed method. In this research the ISBR data bank is used for simmulations.Moreover in medical applications getting patient condition and information with fast methods is very important especially in emergency circumstances. Therefore all effective agents in patient health must be fast even medical algorithms such as clustering ones . Hence in this paper to decrease the time of convergence considerably and decline the number of iterations significantly, cluster centroids are initialized by an algorithm.
Biomedical Image Processing / Medical Image Processing
Ladan Amini; Hamid Soltanian Zadeh; Caro Lucas; Masoume Giti
Volume -2, Issue 1 , July 2005, , Pages 17-34
Abstract
Based on a discrete dynamic contour model, a method for segmentation of brain structures like thalamus and red nucleus from magnetic resonance images (MRI) is developed. A new method for solving common problems in extracting the discontinuous boundary of a structure from a low contrast image is presented. ...
Read More
Based on a discrete dynamic contour model, a method for segmentation of brain structures like thalamus and red nucleus from magnetic resonance images (MRI) is developed. A new method for solving common problems in extracting the discontinuous boundary of a structure from a low contrast image is presented. External and internal forces deform the dynamic contour model. Internal forces are obtained from local geometry of the contour, which consist of vertices and edges, connecting adjacent vertices. The image data and desired image features such as image energy are utilized to obtain external forces. The problem of low contrast image data and unclear edges in the image energy is overcome by the proposed algorithm that uses several methods like thresholding, unsupervised clustering methods such as fuzzy C-means (FCM), edge-finding filters like Prewitt, and morphological operations. We also present a method for generating an initial contour for the model from the image data automatically. Evaluation and validation of the methods are conducted by comparing radiologist and automatic segmentation results. The average of the similarity between segmentation results is 0.8 for the left and right thalami indicating excellent performance of the new method. Additional noise and intensity inhomogeneity changed the evaluation results slightly illustrating the robustness of the proposed method to the image noise and intensity inhomogeneity.