Nikolay Metodiev Sirakov

Lesion detection in demoscopy images with novel density-based and active contour approaches

BackgroundDermoscopy is one of the major imaging modalities used in the diagnosis of melanoma and other pigmented skin lesions. Automated assessment tools for dermoscopy images have become an important field of research mainly because of inter- and intra-observer variations in human interpretation. One of the most important steps in dermoscopy image analysis is the detection of lesion borders, since many other features, such as asymmetry, border irregularity, and abrupt border cutoff, rely on the boundary of the lesion.ResultsTo automate the process of delineating the lesions, we employed Active Contour Model (ACM) and boundary-driven density-based clustering (BD-DBSCAN) algorithms on 50 dermoscopy images, which also have ground truths to be used for quantitative comparison. We have observed that ACM and BD-DBSCAN have the same border error of 6.6% on all images. To address noisy images, BD-DBSCAN can perform better delineation than ACM. However, when used with optimum parameters, ACM outperforms BD-DBSCAN, since ACM has a higher recall ratio.ConclusionWe successfully proposed two new frameworks to delineate suspicious lesions with i) an ACM integrated approach with sharpening and ii) a fast boundary-driven density-based clustering technique. ACM shrinks a curve toward the boundary of the lesion. To guide the evolution, the model employs the exact solution [27] of a specific form of the Geometric Heat Partial Differential Equation [28]. To make ACM advance through noisy images, an improvement of the model’s boundary condition is under consideration. BD-DBSCAN improves regular density-based algorithm to select query points intelligently.

A New Active Convex Hull Model for Image Regions

This paper presents a new active convex hull model with the following advantages: invariant with respect to the number of pixels to be enveloped; the number of time iterations is invariant, with respect to the image size; time-cheap for large image regions. The model is based on the geometric heat differential equations, derived from parabolic equation, and parameterized by arc length. To prevent the active contour from intruding into concavities and evolve it to the proper convex hull we use a vector field given as a difference between normal and tangent forces. The vector field is also used to segment an image to shells, such that a single region is present in each shell. A penalty function is developed to stop evolvement of those arc segments, whose vectors encountered boundary points of an image region. Based on the model a discrete algorithm is designed and coded by Mathematica 5.2. A condition is developed, with respect to the image size, to guarantee stable convergence of the active contour to the convex hull of the desired region. To validate the advantages and contributions a set of experiments is performed using synthetic, groundwater and medical images of different size and modalities. The paper concludes with a discussion and comparison of the active convex hull model with set of existing convex hull algorithms.

Dermoscopic diagnosis of melanoma in a 4D space constructed by active contour extracted features

Dermoscopy, also known as epiluminescence microscopy, is a major imaging technique used in the assessment of melanoma and other diseases of skin. In this study we propose a computer aided method and tools for fast and automated diagnosis of malignant skin lesions using non-linear classifiers. The method consists of three main stages: (1) skin lesion features extraction from images; (2) features measurement and digitization; and (3) skin lesion binary diagnosis (classification), using the extracted features. A shrinking active contour (S-ACES) extracts color regions boundaries, the number of colors, and lesions boundary, which is used to calculate the abrupt boundary. Quantification methods for measurements of asymmetry and abrupt endings in skin lesions are elaborated to approach the second stage of the method. The total dermoscopy score (TDS) formula of the ABCD rule is modeled as linear support vector machines (SVM). Further a polynomial SVM classifier is developed. To validate the proposed framework a dataset of 64 lesion images were selected from a collection with a ground truth. The lesions were classified as benign or malignant by the TDS based model and the SVM polynomial classifier. Comparing the results, we showed that the latter model has a better f-measure then the TDS-based model (linear classifier) in the classification of skin lesions into two groups, malignant and benign.

An Integral Active Contour Model for Convex Hull and Boundary Extraction

This paper presents a new deformable model capable of segmenting images with multiple complex objects and deep concavities. The method integrates a shell algorithm, an active contour model and two active convex hull models. The shell algorithm automatically inscribes every image object into a single convex curve. Every curve is evolved to the boundarys vicinity by the exact solution of a specific form of the heat equation. Further, if re-parametrization is applied at every time step of the evolution the active contour will converge to deep concavities. But if distance function minimization or line equation is used to stop the evolution, the active contour will define the convex hull of the object. Set of experiments are performed to validate the theory. The contributions, the advantages and bottlenecks of the model are underlined at the end by a comparison against other methods in the field.

Automatic boundary detection and symmetry calculation in dermoscopy images of skin lesions

This paper develops an approach and tool that automatically extracts skin lesions boundary used for symmetry and area calculation. An image enhancement approach prepares every image for active contour (AC) evolution. Further, the AC automatically extracts the lesions boundary used to measure symmetry applying minimal boundary box. Next, the lesions area is calculated. Thus, the lesions are mapped as points onto area - symmetry 2D space to determine the distribution of the lesions with cancer. To validate the theoretical concepts experiments were performed with 51 skin lesion images. A statistics measures the accuracy of boundary extraction with respect to a ground truth. The advantages, disadvantages and the contribution of this study are reported at the end of the paper.

Recognition of Emotional states in Natural Human-Computer Interaction

In this paper we present a non-invasive method for extracting facial expression components from video sequences. We propose a contextual analysis of user state and guide appropriate system actuation. The approach proposed hereafter combines the advantages of MPEG-4 and an active contour model to extract the contours of the facial objects such as: eyes, eyebrows, nose, lips and dimples. The first stage applies a local statistics to distinguish the objects subject of interest. The second stage runs an existing active contour model to define the contours of the objects. Further the facial control points could be spotted on the contours and used for emotion recognition. To distinguish geometric facial features an approach is proposed to compare multiple closed polygons. To validate the theoretical concepts experiments were performed using a normal and a smiling face. A comparison with an existing approach underlines the advantages and disadvantages of the present work.

Optimal set of features for accurate skin cancer diagnosis

Skin cancer is on the rise. Hence the accurate detection of cancerous lesions is of paramount importance in the treatment of this health condition. In this study, we present a computer vision framework that studies 10 skin lesion features including newly introduced morphological and texture features along with established in the literature. All features are extracted automatically from 90 lesion images. A two-class classification problem was applied to determine the most significant features of disease using three features selection methods, Support Vector Machines Recursive Feature Elimination (SVMRFE), Information Gain, and Correlation-based Feature Subset Selection. We found that the five features selected by SVMRFE provides the highest accuracy readings of 100% model, 84% leave-one-out, and 89% 10-fold cross validation (10×CV) than the other two methods. Comparing the selected features with those used by the Total Dermoscopy Score, we report consistencies, disagreements, and the contributions of the present work.

Efficient segmentation with the convex local-global fuzzy Gaussian distribution active contour for medical applications

A new active contour (LGFGD) was developed in our earlier conference paper. This contour uses local and global information along with Gaussian distribution. The present paper derives the main LGFGD equation and investigates its parameters σ, λ and m. Specific values are determined (for σ, λ, m) to ensure high accuracy of segmentation of medical images containing nonhomogeneous and noisy regions with week boundaries. To validate the model, a new set of experiments was performed with new images including 24 skin lesion images with ground truth. Thus, a statistic of the LGFGD performance was calculated regarding the model’s interval of confidence. Comparison with contemporary methods from the field is provided as well.

Threat assessment using visual hierarchy and conceptual firearms ontology

Abstract. The work that established and explored the links between visual hierarchy and conceptual ontology of firearms for the purpose of threat assessment is continued. The previous study used geometrical information to find a target in the visual hierarchy and through the links with the conceptual ontology to derive high-level information that was used to assess a potential threat. Multiple improvements and new contributions are reported. The theoretical basis of the geometric feature extraction method was improved in terms of accuracy. The sample space used for validations is expanded from 31 to 153 firearms. Thus, a new larger and more accurate sequence of visual hierarchies was generated using a modified Gonzalez’ clustering algorithm. The conceptual ontology is elaborated as well and more links were created between the two kinds of hierarchies (visual and conceptual). The threat assessment equation is refined around ammunition-related properties and uses high-level information from the conceptual hierarchy. The experiments performed on weapons identification and threat assessment showed that our system recognized 100% of the cases if a weapon already belongs to the ontology and in 90.8% of the cases, determined the correct third ancestor (level concept) if the weapon is unknown to the ontology. To validate the accuracy of identification for a very large data set, we calculated the intervals of confidence for our system.

Skin lesion feature vectors classification in models of a Riemannian manifold

This study is a continuation of a work published by Mete, Ou, and Sirakov (2012), where a model of a 4D manifold of feature vectors was developed. The present paper introduces an improved metric in the 4D manifold first and then extends both the size of the sample space and the dimension (to 6D) of the manifold model in which the sample space lies. As a result, we not only overcame the issue of one single vector representing multiple skin lesions, which occurred in the work of Mete, Ou, and Sirakov (2012), but also improved the accuracy of classification. Furthermore, a statistical evaluation of our support vector machine (SVM) classification method was performed. The intervals of confidence were calculated for the mean of classification of a large sample set in the 6D model. Comparison results of classification with our SVM in 4D and 6D models using 10-fold cross-validation are given at the end of the paper. It is found that the 6D model improves the classification results of the previous study suggesting that two newly introduced features contributed to the increase of the classification accuracy.