Marcos Vinicius Naves Bedo

Endowing a Content-Based Medical Image Retrieval System with Perceptual Similarity Using Ensemble Strategy

Content-based medical image retrieval (CBMIR) is a powerful resource to improve differential computer-aided diagnosis. The major problem with CBMIR applications is the semantic gap, a situation in which the system does not follow the users’ sense of similarity. This gap can be bridged by the adequate modeling of similarity queries, which ultimately depends on the combination of feature extractor methods and distance functions. In this study, such combinations are referred to as perceptual parameters, as they impact on how images are compared. In a CBMIR, the perceptual parameters must be manually set by the users, which imposes a heavy burden on the specialists; otherwise, the system will follow a predefined sense of similarity. This paper presents a novel approach to endow a CBMIR with a proper sense of similarity, in which the system defines the perceptual parameter depending on the query element. The method employs ensemble strategy, where an extreme learning machine acts as a meta-learner and identifies the most suitable perceptual parameter according to a given query image. This parameter defines the search space for the similarity query that retrieves the most similar images. An instance-based learning classifier labels the query image following the query result set. As the concept implementation, we integrated the approach into a mammogram CBMIR. For each query image, the resulting tool provided a complete second opinion, including lesion class, system certainty degree, and set of most similar images. Extensive experiments on a large mammogram dataset showed that our proposal achieved a hit ratio up to 10% higher than the traditional CBMIR approach without requiring external parameters from the users. Our database-driven solution was also up to 25% faster than content retrieval traditional approaches.

A Label-Scaled Similarity Measure for Content-Based Image Retrieval

Content-Based Image Retrieval (CBIR) has proven to be a suitable complement to traditional text-based searching. CBIR applications rely on two main steps, namely the representation of the images, and the similarity measuring between two represented images. Although modern segmentation and learning algorithms enable the accurate representation of local and global features within an image, how to properly compare the segmented objects is still an open issue. In this study, we propose a new comparison method called Counting-Labels Similarity Measure (CL-Measure). Our approach calculates the similarity between two images by comparing the labeled regions within these images and by balancing the influence of each label according to its predominance in both non-metric and metric fashion. The experiments on a real dataset of dermatological ulcers show that CL-Measure achieves a higher Precision for all values of Recall compared to its competitors in retrieval tasks.

Unveiling smoke in social images with the SmokeBlock approach

Can we use information from social media and crowdsourced images to detect smoke and assist rescue forces? While there are computer vision methods for detecting smoke, they require movement information extracted from video data. In this paper we propose SmokeBlock: a method that is able to segment and detect smoke in still images. SmokeBlock uses superpixel segmentation and extracts local color and texture features from images to spot smoke. We used real data from Flickr and compared SmokeBlock against state-of-the-art methods for feature extraction. Our method achieved performance superior than the competitors, for the task of smoke detection. Our findings shall support further investigations in the field of image analysis, in particular, concerning images captured with mobile devices.

Color and Texture Influence on Computer-Aided Diagnosis of Dermatological Ulcers

This study presents an analysis of classification techniques for Computer-Aided Diagnosis (CAD) regarding ulcerated lesions. We focus on determining influence of both color and texture in the automated image classification and its implication. To do so, we assayed a dataset of dermatological ulcers containing five variations in terms of tissue composition of lesion skin: granulation (red), fibrin (yellow), callous (white), necrotic (black), and a mix of the previous variations (mixed). Every image was previously labelled by experts regarding this red-yellow-black-white-mixed model. We employed specially designed color and texture extractors to represent the dataset images, namely: Color Layout, Color Structure, Scalable Color, Edge Histogram, Haralick, and Texture-Spectrum. The first three are color feature extractors and the last three are texture extractors. Following, we employed the Symmetrica Uncert Attribute Eval method to determine the features suitable for image classification. We tested a set of classifiers that follows distinct paradigms over the selected features, achieving an accuracy ratio of up to 77% in terms of images correctly classified, with the area under the receiver operating characteristic (ROC) curve up to 0.84. The classification performance and the selected features enabled us to determine that texture features were more predominant than color in the entire classification process.

A Similarity-Based Approach for Financial Time Series Analysis and Forecasting

Financial time series analysis have been attracting research interest for several years. Many works have been proposed to perform economic series forecasting, however, it still is a hard endeavor to develop a general model that is able to handle the chaotic nature of the markets. Artificial intelligence methods such as artificial neural networks and support vector machines arose as promising alternatives, but they hide the processing semantics, limiting the result interpretation. In addition, one of the main drawbacks of the existing solutions is that they usually cannot be easily employed as building blocks of new analysis tools. This paper presents a new approach to financial time series forecasting based on similarity between series patterns using a database-driven architecture. We propose a new feature extractor based on visual features associated with a boosted instance-based learning classifier to predict a share’s behavior, thus improving the human analyst understanding and validation of the results. The analysis is defined through extended SQL instructions and executed over a fast and scalable engine, which makes our solution adequate to provide data analysis support for new applications handling large time series datasets. We also present experiments performed on data obtained from distinct market shares. The achieved results show that our approach outperformed existing methods in terms of accuracy, running time and scalability.

The Merkurion approach for similarity searching optimization in Database Management Systems

Modern Database Management Systems (DBMSs) retrieve songs that resemble those in a music dataset, identify plagiarism in a set of documents, or provide past cases to physicians by taking into account the characteristics of a query exam. All such tasks require the comparison of data by similarity, which can be expressed in terms of distance-based queries in metric spaces. Traditional query processing relies mostly on histograms for describing the data distribution space and choosing a data retrieval path that quickly leads to the answer, discarding comparisons of most unwanted data. However, DBMSs still lack adequate support for selectivity estimation of query operators for data types embedded in metric spaces. This article addresses a novel strategy that extends the query optimizer of any DBMS, so that it can also perform both logical and physical query plan optimizations in searches that include similarity predicates. The proposal, named Merkurion, updates the concept of Data Distribution Space and captures data distributions according to the distances between the elements within a dataset. Moreover, it employs concise representations of such distributions, called synopses, for the definition of rules that enable similarity searching optimization. An extensive evaluation of Merkurion in real-world datasets has proven its effectiveness and broad applicability to many data domains.

When Similarity is Not Enough, Ask for Diversity: Grouping Elements Based on Influence

Crowdsourcing images have been increasingly employed for mapping emergency scenarios, which helps rescue forces in choosing contingency plans. In this scenario, similarity searching can be used to retrieve related images from past situations. However, the retrieved images often are similar among themselves and, therefore, add little to none new information to the rescue decision-making process. In this paper, we take advantage of diversity queries to increase the variety of the representative elements about an incident, whereas the remaining and related data are grouped according to the set of representatives. Thus, our approach enables content retrieval, grouping and an easier exploration of the result set. Experiments performed on real datasets shows that our proposal outperforms the existing methods regarding both quality and performance, being at least three orders of magnitude faster.

Fire Detection from Social Media Images by Means of Instance-Based Learning

Social media can provide valuable information to support decision making in crisis management, such as in accidents, explosions, and fires. However, much of the data from social media are images, which are uploaded at a rate that makes it impossible for human beings to analyze them. To cope with that problem, we design and implement a database-driven architecture for fast and accurate fire detection named FFireDt. The design of FFireDt uses the instance-based learning through indexed similarity queries expressed as an extension of the relational Structured Query Language. Our contributions are: (i) the design of the Fast-Fire Detection (\(FFireDt\)), which achieves efficiency and efficacy rates that rival to the state-of-the-art techniques; (ii) the sound evaluation of 36 image descriptors, for the task of image classification in social media; (iii) the evaluation of content-based indexing with respect to the construction of instance-based classification systems; and (iv) the curation of a ground-truth annotated dataset of fire images from social media. Using real data from Flickr, the experiments showed that system \(FFireDt\) was able to achieve a precision for fire detection comparable to that of human annotators. Our results are promising for the engineering of systems to monitor images uploaded to social media services.

Being Similar is Not Enough: How to Bridge Usability Gap through Diversity in Medical Images

In this paper we present a technique developed to bridge the usability gap in Content-Based Medical Image Retrieval (CBMIR) systems exploring both similarity and diversity. Usability gaps are related to how easy to use a software tool from the radiologists perspective is. Although much have been done to better express similarity queries, the use of CBMIR over massive databases may have drawbacks that impact its usability. We claim that much of the problems derives from the fact that many images returned are closer to each other than to the query element (near-duplicates). To target this nuisance, we propose to boost similarity queries with diversity, using a technique to hierarchically cluster near-duplicates. We tailored a domain-independent and parameter-free method by controlling the maximum area reached in the search space. This novel approach to improve CBMIR systems take advantage of diversity expectations. The proposed approach BridGE (Better result with influence diversification to Group Elements) aims at adding new relevant information to the analysts, reducing the need of further query refinement or relevance feedback cycles. The results are displayed to the specialist as a traditional CBMIR result whereas the radiologists are able to expand the clusters and navigate through them. The results support our claim that a CBMIR system empowered with diversity is able to bridge the usability gap, grouping near-duplicates and being at least 2 orders of magnitude faster than its mainly competitors.