Pradeep Chowriappa

Wavelet-Based Energy Features for Glaucomatous Image Classification

Texture features within images are actively pursued for accurate and efficient glaucoma classification. Energy distribution over wavelet subbands is applied to find these important texture features. In this paper, we investigate the discriminatory potential of wavelet features obtained from the daubechies (db3), symlets (sym3), and biorthogonal (bio3.3, bio3.5, and bio3.7) wavelet filters. We propose a novel technique to extract energy signatures obtained using 2-D discrete wavelet transform, and subject these signatures to different feature ranking and feature selection strategies. We have gauged the effectiveness of the resultant ranked and selected subsets of features using a support vector machine, sequential minimal optimization, random forest, and naïve Bayes classification strategies. We observed an accuracy of around 93% using tenfold cross validations to demonstrate the effectiveness of these methods.

Thyroid lesion classification in 242 patient population using Gabor transform features from high resolution ultrasound images

Total of 242 benign and malignant thyroid nodules are classified.Various entropies are extracted from Gabor transformed images.These features are subjected to LSDA and ranked by Relief-F method.Various sampling strategies are used to balance the classification data.Obtained classification accuracy of 94.3% with C4.5 decision tree classifier. Thyroid cancer commences from an atypical growth of thyroid tissue at the edge of the thyroid gland. Initially, it forms a lump in the throat and an over-growth of this tissue leads to the formation of benign or malignant thyroid nodules. Blood test and biopsies are the standard techniques used to diagnose the presence of thyroid nodules. But imaging modalities can improve the diagnosis and are marked as cost-effective, non-invasive and risk-free to identify the stages of thyroid cancer. This study proposes a novel automated system for classification of benign and malignant thyroid nodules. Raw images of thyroid nodules recorded using high resolution ultrasound (HRUS) are subjected to Gabor transform. Various entropy features are extracted from these transformed images and these features are reduced by locality sensitive discriminant analysis (LSDA) and ranked by Relief-F method. Over-sampling strategies with Wilcoxon signed-rank, Friedmans and Iman-Davenport post hoc tests are used to balance the classification data and also to improve the classification performance. Classifiers such as support vector machine (SVM), k-nearest neighbour (kNN), multi-layered perceptron (MLP) and decision tree are used for the characterization of benign and malignant thyroid nodules. We have obtained a classification accuracy of 94.3% with C4.5 decision tree classifier using 242 thyroid HRUS images. Our developed system can be used to screen the thyroid automatically and assist the radiologists.

Ensemble selection for feature-based classification of diabetic maculopathy images

As diabetic maculopathy (DM) is a prevalent cause of blindness in the world, it is increasingly important to use automated techniques for the early detection of the disease. In this paper, we propose a decision system to classify DM fundus images into normal, clinically significant macular edema (CMSE), and non-clinically significant macular edema (non-CMSE) classes. The objective of the proposed decision system is three fold namely, to automatically extract textural features (both region specific and global), to effectively choose subset of discriminatory features, and to classify DM fundus images to their corresponding class of disease severity. The system uses a gamut of textural features and an ensemble classifier derived from four popular classifiers such as the hidden naïve Bayes, naïve Bayes, sequential minimal optimization (SMO), and the tree-based J48 classifiers. We achieved an average classification accuracy of 96.7% using five-fold cross validation.

Protein Structure Classification Based on Conserved Hydrophobic Residues

Protein folding is frequently guided by local residue interactions that form clusters in the protein core. The interactions between residue clusters serve as potential nucleation sites in the folding process. Evidence postulates that the residue interactions are governed by the hydrophobic propensities that the residues possess. An array of hydrophobicity scales has been developed to determine the hydrophobic propensities of residues under different environmental conditions. In this work, we propose a graph-theory-based data mining framework to extract and isolate protein structural features that sustain invariance in evolutionary-related proteins, through the integrated analysis of five well-known hydrophobicity scales over the 3D structure of proteins. We hypothesize that proteins of the same homology contain conserved hydrophobic residues and exhibit analogous residue interaction patterns in the folded state. The results obtained demonstrate that discriminatory residue interaction patterns shared among proteins of the same family can be employed for both the structural and the functional annotation of proteins. We obtained on the average 90 percent accuracy in protein classification with a significantly small feature vector compared to previous results in the area. This work presents an elaborate study, as well as validation evidence, to illustrate the efficacy of the method and the correctness of results reported.

Introduction to Machine Learning in Healthcare Informatics

Healthcare informatics, a multi-disciplinary field has become synonymous with the technological advancements and big data challenges. With the need to reduce healthcare costs and the movement towards personalized healthcare, the healthcare industry faces changes in three core areas namely, electronic record management, data integration, and computer aided diagnoses. Machine learning a complex field in itself offers a wide range of tools, techniques, and frameworks that can be exploited to address these challenges. This chapter elaborates on the intricacies of data handling the data rich filed of healthcare informatics, and the potential role of machine learning to mitigate the challenges faced.

Region Quad-Tree Decomposition Based Edge Detection for Medical Images

Edge detection in medical images has generated significant interest in the medical informatics community, especially in recent years. With the advent of imaging technology in biomedical and clinical domains, the growth in medical digital images has exceeded our capacity to analyze and store them for efficient representation and retrieval, especially for data mining applications. Medical decision support applications frequently demand the ability to identify and locate sharp discontinuities in an image for feature extraction and interpretation of image content, which can then be exploited for decision support analysis. However, due to the inherent high dimensional nature of the image content and the presence of ill-defined edges, edge detection using classical procedures is difficult, if not impossible, for sensitive and specific medical informatics-based discovery. In this paper, we propose a new edge detection technique based on the regional recursive hierarchical decomposition using quadtree and post-filtration of edges using a finite difference operator. We show that in medical images of common origin, focal and/or penumbral blurred edges can be characterized by an estimable intensity gradient. This gradient can further be used for dismissing false alarms. A detailed validation and comparison with related works on diabetic retinopathy images and CT scan images show that the proposed approach is efficient and accurate.

A framework to identify influencers in signed social networks

Social networks are defined as a graphical data structure, which captures complex social interactions between users of a social network. Signed social networks are weighted representations of the social network with the emphasis of capturing both positive and negative interactions (edges) between actors of the network. Ad-hoc communities in a social network, as a corollary can be treated as the logical grouping of social actors that share common interests, ideas, or beliefs. In this work, we leverage these known constructs in social networks to effectively identify influencers (i.e. a subset of actors that exert their influence over a community), aka, seeds. Traditional approaches largely rely on degree of connectivity in identifying influencers of a community. We hypothesize that there are other measures to identify influences. In this work, our objective is therefore to explore and propose a technique using Principal Component Analysis (PCA) to identify the smallest set of influencers with increasing the possibility of adopting a product. Furthermore, we validate our finding by evaluating the potential of these influencers to identify positive communities in a social network. We believe our approach is novel in choosing our influencers (seeds) and thus by using these seeds, positive and negative edges are established. We exploit resulting positive and negative edges to mine ad-hoc communities of interest.

Capturing hydrophobic moment using spectral coherence for protein secondary structure prediction

The endeavor to decipher the structure and function of a protein from its amino acid sequence has provided an enduringly interesting challenge. Due to the sheer quantity of existing protein data, this challenge naturally presents itself as a complex computational problem requiring the deployment of novel data mining techniques. We hypothesize that the hydrophobic moment (HM) is important in the folding and the formation of secondary structures, and is evolutionarily retained in structurally related proteins. We propose the use of magnitude-squared spectral coherence (MSC) to capture HM of a sequence using selected hydrophobicity scales for effective structural and fold classification of protein sequences. Extensive experimentation on PDBselect dataset demonstrates overall success rates of 77.4% and 63.4% for structural and fold classification. The comparative results show that spectral coherence between the hydrophobic and hydrophilic representations of a sequence effectively captures periodic hydrophobic variations over the length of the sequence that corresponds to HM.

Ranking through integration of protein-similarity for identification of cell-cyclic genes

Gene array experiments are progressively conducted. However, the biological functional interpretation has not kept pace with this rapid escalation. Functional genomics using data mining methods potentially offers precise, objective, and more reliable gene identification. Our work creates a gene-ranking scheme by integrating gene expression profile phase information with protein similarity to identify cell-cyclic genes. We present a unique schema to enable integration by employing QR-factorisation from the pair-wise similarity matrix formulation. Angular coefficients are derived and consequently employed for integrated gene ranking. Experimental results on an independent benchmark dataset signify the efficacy of the method.

Multi-domain Protein Family Classification Using Isomorphic Inter-property Relationships

Multi-domain proteins result from the duplication and combination of complex but limited number of domains. The ability to distinguish multi-domain homologs from unrelated pairs that share a domain is essential to genomic analysis. Heuristics based on sequence similarity and alignment coverage have been proposed to screen out domain insertions but have met with limited success. In this paper we propose a unique protein classification schema for multi-domain protein superfamilies. Segmented profiles of physico-chemical properties and amino acid composition are created for vector quantization based dimensionality reduction to create a feature profile for rule-discovery and classification. Association rules are mined to identify isomorphic relationships that govern the formation of domains between proteins to correctly predict homologous pairs and reject unrelated pairs, including those that share domains. Our results demonstrate that effective classification of conserved domain classes can be performed using these feature profiles, and the classifier is not susceptible to class imbalances frequently encountered in these databases.