%0 Thesis %A Jalalian, Afsaneh %D 2017 %G English %T Computer-assisted diagnosis system for angiogenesis detection and classification in Computed Tomography Laser Mammography %U http://ethesis.upm.edu.my/id/eprint/12505/1/FK%202017%2067%20-%20T.pdf %X Computed Tomography Laser Mammography (CTLM) is a full tomographic system to explore neo-angiogenesis in the breast by generating a volumetric image. Angiogenesis is a new forming of blood vessels which supply the tumour and seen in different shapes such as free standing, polypoid, ring shaped, dumb-bell shaped, diverticular, and spindle shaped. The manual detection of angiogenesis and differentiation of the shapes is a challenging procedure for physician and a CAD system is expected to help radiologists as a second reader. In this research, a CAD on CTLM images is proposed to detect and classify the angiogenesis. The proposed CAD systems contain four main steps which are segmentation, objects dissociation, feature extraction, and classification. In segmentation stage, three automatic segmentation techniques are implemented to extract and reconstruct the volume of interests (VOIs) on CTLM images. The ground truth is extracted from window-level technique on the original CTLM images. As the pre-processing before feature extraction step, the VOIs have been dissociated to sub-VOIs. The two region properties features include centroid and extrema which have been utilised to prepare the dissociation model of VOIs. According to the characteristics of abnormalities in CTLM image that critically depends on shape and intensity, various shape and texture properties are extracted in the feature extraction level. Three different compactness features are extracted from dissociated objects (sub- VOIs). The Harlick’s features are extracted based on 3D Grey Level Co-occurrence (GLCM) matrix. Hence, different combination of shape features and Harlick’s features have been used for the training procedure. In the image classification, support vector machine (SVM) and multilayer perceptron neural network (MLPNN) have been used to classify the abnormality in CTLM images. CTLM data set in this work includes 180 patients which are diagnosed by two expert radiologists that considered 132 cases as benign and 48 cases as malignant. In order to overcome the imbalanced dataset, various techniques such as soft margin, kernel function transformation, and oversampling method have been applied to enhance the performance of the proposed classifiers. The Jaccard and Dice coefficients in addition to the volumetric overlap error are employed to quantify the accuracy of segmentation methods. According to the outcomes, the 3D Fuzzy C-Means clustering presents reasonable results compared to other methods. The K-fold cross-validation with k=10 is used in the training and test of the proposed classifier. The experimental results show that SVM with radial basis function (SVMRBF) on oversampled data by Adaptive Synthetic Sampling (ADASYN) method achieved the highest performance in terms of accuracy, sensitivity, and specificity which are 98.6%, 97.78% and 99.43%, respectively. The results of angiogenesis diagnosis by SVM-RBF on oversampled data by ADASYN completely matched with the reports of two expert radiologists in localisation and shapes of angiogenesis. The proposed CTLM-CAD recognise the diverticular shape, polypoid, spindle shaped and free standing shape of angiogenesis.