Sujatha Cm

Analysis of ventricle regions in Alzheimer’s brain MR images using level set based methods

In this work, an attempt has been made to analyse ventricle region of the T1 weighted coronal Magnetic Resonance (MR) brain images and study the progression of severity in Alzheimer’s Disease (AD) conditions. Two level set methods namely Distance Regularised Level Set Evolution (DRLSE) and geodesic active contour are used to extract the desired region of interest. Eighty geometric features are derived from the segmented ventricle region. The most significant parameters are found using principal component analysis. Results demonstrate that the DRLSE shows better performance in extraction of the boundary of the ventricle region than geodesic active contour method. The geometrical feature, area is found to have a high correlation with brain to ventricle index for all subjects. Further, it is observed that this feature gives a distinct separation between normal and abnormal AD subjects (p value = 0.00012). It also provides high correlation for normal (.97) and abnormal AD subjects (<0.9). Hence, this analysis could be a useful supplement to physicians in diagnosis and treatment of Alzheimer’s and other neurodegenerative disorders.

A Method to Differentiate Mild Cognitive Impairment and Alzheimer in MR Images using Eigen Value Descriptors

Automated analysis and differentiation of mild cognitive impairment and Alzheimer’s condition using MR images is clinically significant in dementic disorder. Alzheimer’s Disease (AD) is a fatal and common form of dementia that progressively affects the patients. Shape descriptors could better differentiate the morphological alterations of brain structures and aid in the development of prospective disease modifying therapies. Ventricle enlargement is considered as a significant biomarker in the AD diagnosis. In this work, a method has been proposed to differentiate MCI from the healthy normal and AD subjects using Laplace-Beltrami (LB) eigen value shape descriptors. Prior to this, Reaction Diffusion (RD) level set is used to segment the ventricles in MR images and the results are validated against the Ground Truth (GT). LB eigen values are infinite series of spectrum that describes the intrinsic geometry of objects. Most significant LB shape descriptors are identified and their performance is analysed using linear Support Vector Machine (SVM) classifier. Results show that, the RD level set is able to segment the ventricles. The segmented ventricles are found to have high correlation with GT. The eigen values in the LB spectrum could show distinction in the feature space better than the geometric features. High accuracy is observed in the classification results of linear SVM. The proposed automated system is able to distinctly separate the MCI from normal and AD subjects. Thus this pipeline of work seems to be clinically significant in the automated analysis of dementic subjects.

Multilevel Tsallis entropy based segmentation for detection of object and shadow in SONAR images

In the proposed work, Multilevel Tsallis entropy is employed to segment side scan SONAR images. Bilevel thresholding detects only the object whereas Multilevel thresholding detects both object and its shadow in a given image. Shadow detection provides more details of the target [1]. In Tsallis entropy method, initially an objective function is evaluated. This function considers the probability density function of the pixels distribution to evaluate the threshold for bilevel and multilevel. Objective function is a function of entropic index `q and is varied from 0.1 to 0.9 to determine the threshold value for bilevel. For obtaining multilevel threshold, `q is varied from 2 to 9. PSNR is calculated for both bilevel and multilevel to select the optimal threshold. Based on the threshold value, image is segmented and analysed. The results show that Tsallis multilevel thresholding is able to segment the object and the shadow in all the considered images. PSNR values for the considered images are 15 dB, 15.1 dB, 17.9 dB for bilevel 15.8 dB, 17.5 dB and 18 dB for multilevel thresholding.

Swarm intelligence based segmentation for buried object scanning SONAR images

In the present work, Particle swarm optimisation (PSO) based Tsallis entropy method is employed to segment the buried object SONAR images. This SONAR detects the objects present beneath the seabed in ocean. Objects may be pipelines, and unexploded ordinances buried beneath the seabed. Computer vision for object detection is required when SONAR is equipped in autonomous underwater vehicle. The vehicle acquires volumes of data to be analysed manually which is time consuming and expensive. In Tsallis entropy segmentation method, PSO based optimisation technique is employed to select the appropriate bilevel thresholds for every image. For the considered image `mild steel and concrete threshold value is 129 and the corresponding accuracy is 99.14 %. The threshold value for the image `stone is 132 and the accuracy is 97.5 % for `q value 0.2.

Analysis of Alzheimer MR brain images using entropy based segmentation and Minkowski Functional

In this work, an attempt has been made to analyze atrophy of MR brain images using Minkowski Functionals (MFs) of the entropy based skull stripped whole brain image. The normal and Alzheimer images considered in this work are obtained from MIRIAD database. The proposed algorithm uses Shannon entropy and Tsallis entropy methods to calculate the global and local threshold values for the edge detection. The obtained edges map are further processed using morphological operation. The mask generated from the edge map is used to extract the brain tissues. The performance of skull stripping is validated by correlating the total brain area and ground truth. The accuracy of entropy based skull stripping is compared with Otsu thresholding method. The structural changes in skull stripped brain images are analysed using Minkowski functionals such as area, perimeter and Euler number. Results show that the entropy based method is able to extract the total brain. The correlation of total brain area with ground truth is high (R = 0.93). It is also observed that the Minkowski functional, Euler number gives significant discrimination (p<;0.001) of normal and Alzheimer subjects. Hence, the entropy based method along with Minkowski functionals could be used for diagnosis of Alzheimer conditions in the brain.

Particle Swarm Optimization-Based SONAR Image Enhancement for Underwater Target Detection

In the proposed work, particle swarm optimization (PSO) is employed to enhance SONAR images in spatial domain. A transformation function is used with the local and global data of the image to enhance the image based on PSO. The objective function considers the entropy of the image, number of edges detected, and the sum of edge intensities. PSO determines the optimal parameters required for image enhancement. PSO-based image enhancement is compared with median filter and adaptive Wiener filter both qualitatively and quantitatively. Mean square error (MSE) and peak signal-to-noise ratio (PSNR) are used as quantitative measures for the analysis of the enhanced images. MSE and PSNR analysis reveals that original details of the image are retained after enhancement. Based on the visual and quantitative analysis, it is considered that PSO-based technique provides better enhancement of SONAR images.

Prediction of FEV 1 and FEV 6 in normal and obstructive abnormality using ELM regression and spirometric investigations

Spirometric pulmonary function test is a well-established test in clinical medicine for the assessment of respiratory diseases. It measures the volume of air inhaled or exhaled as a function of time during forced breathing maneuvers and generates large data set. However, spirometric investigation is often prone to incomplete data sets due to inability of the children and patient to perform this test. Hence, there is a requirement for prediction of significant parameters from incomplete data set. In this work, Extreme Learning Machines (ELM) based prediction of significant parameters which include forced expiratory volumes at first second and sixth seconds are employed. ELM with radial basis function and sine activation functions is used to train and test the prediction process. The performance of the regression model is validated with prediction accuracy and root mean squared error. Results demonstrate that ELM with Radial Basis Function (RBF) achieves higher prediction accuracy (98%) for FEV1 prediction with 20 hidden neurons. Prediction accuracy of 95% is obtained for FEV6 with sine activation function. Further, it is observed that less prediction error and high correlation are obtained in FEV1 prediction. It appears that this method of prediction of significant parameter could be clinically relevant in the diagnosis of pulmonary abnormalities with incomplete data set and missing values.

Analysis of ventricles in alzheimer MR images using coherence enhancing diffusion filter and level set method

Alzheimers disease is a fatal neurodegenerative disorder that causes enlargement of ventricles and shrinkage of whole brain leading to slow progression of dementia. Ventricle enlargement is an important structural biomarker for the diagnosis of AD. Magnetic resonance imaging is a useful non invasive tool which depicts the pathology of brain structures and improves the diagnosis of AD. In this work, an attempt has been made to segment ventricles from both normal and AD MR images using modified distance regularized level set method. The image gradient of coherence enhanced diffusion filtered image is used as an edge map for level set evolution. Geometric features are extracted from the segmented ventricle to identify its clinical significance in the AD diagnosis. Results show that, the proposed method of integration of non linear diffusion filter with level set method is able to segment the ventricles in both normal and AD images. The extracted shape based features are able to differentiate the AD subjects from normals with high statistical significance (p<;0.0001). Thus this study seems to be clinically useful.

ELM Based Classification and Analysis of Spirometric Pulmonary Function Data

Spirometric pulmonary function test plays a critical role in the diagnosis of respiratory diseases and its clinical utility depends on the accuracy of the measured values. Classification of human respiratory functions experimentally recorded with spirometric pulmonary function test is analyzed using Extreme Learning Machines (ELM). In this study, data (N=300) are obtained using gold standard Spirolab II spirometer under controlled protocol from normal, obstructive and restrictive subjects. The considered parameters include Forced Expiratory Volume in first second (FEV1), forced expiratory volume in sixth second, forced vital capacity and peak expiratory flow. The demographic parameters are also considered. These data are analyzed using extreme learning machines with five different basis functions which include sigmoid, sine, hard limit, triangular basis and radial basis functions. Results show that ELM is able to differentiate normal and abnormal subjects. The accuracy, sensitivity and specificity of ELM are found to be high for ELM with sine basis function. The clinical useful index FEV1 is further validated by correlating it with the performance measures of ELM classifier. The methodology, results and discussion are also presented in this study. As automated analysis of spirometric data is significant for classification of prognosis of respiratory diseases, this study seems to be clinical relevant.