Deep Learning has become increasingly frequent in the studies and analysis of medical images. Advances relevant to this area of research improve computer-aided diagnostic systems and help physicians' routine when providing a second opinion. Breast cancer is one of the types most common cancer among women worldwide. Early diagnosis of breast cancer can facilitate treatment and help saves lives. Mammography is the most widely used exam in the clinical routine to diagnose breast cancer. The analysis of the mammogram requires a specialist with experience in medical imaging. Deep Learning and Machine Learning techniques can collaborate computationally with this task. Adapting the hyperparameters provided to deep learning architectures helps improve the results in analyzing and classifying mammogram images. This paper presents a deep learning-based approach to classifying mammogram image regions of interest (ROIs). This approach includes transfer learning, hyperparameter and fine-tuning, and an ensemble with the models that showed the best results. The process demonstrated promising results, with the ensemble reaching 92% accuracy in the classification of mammogram ROIs of the test set and the area under the curve (AUC) value of 0.97 for the best model.

Breast cancer is one of the most common types of cancer among women in the world. The early diagnosis of the anomaly is a priority for more effective and less costly treatment. Mammography is the most used medical exam in diagnosing breast cancer, and the number of images generated grows exponentially. Accurately interpreting and diagnosing mammography images requires that the specialist has extensive experience to perform this task. The manual image classification and the retrieval of similar cases is an expensive and time-consuming process. The historical data in these exams are an excellent source of knowledge and should be considered when analyzing new cases. Thereby, this research aims to improve the classification and semantic annotation of medical images. We propose a method to extract the low-level image features, metadata information, and medical reports and combine them with the high-level information present in a domain ontology. Also, providing an appropriate semantic structure for integrating information enables the classification and semantic image annotation through a reasoner. Thus, this work seeks to facilitate radiologists' and physicians' daily lives involved in breast cancer diagnosis, collaborating with a second opinion for the medical diagnosis. The experiments have shown that the proposed method is feasible and can generate promising results, achieving a 71% sensitivity rate in identifying circumscribed findings. Moreover, it favors classification and diagnosis as it allows semantically annotating images with concepts of multiple abstraction levels. Therefore, semantic annotation reduces the semantic gap and can help to improve CAD and CBIR systems. The semantic annotation also collaborates with medical education to support the teaching and training of reading medical images and medical diagnosis, supporting decision making by providing a second opinion.

Multiple descriptors are employed to represent images in Content-Based Image Retrieval (CBIR) systems. Each descriptor consists of a feature extractor associated with a distance function. An extractor is generally suitable for representing a specific subset of images on a database. The boundary conditions are information used to detect this subset. The use of visualization in CBIR helps to represent the similarity relationship between images, improving the user's understanding of the CBIR system, allowing them to modify parameters to obtain better results. It is proven that the use of multiple descriptors with boundary conditions tends to improve the precision of CBIR queries, but there is no data on the impact that the technique generates on visualization. This paper uses multiple descriptors with boundary conditions to generate a Neighbor Joining similarity tree-based view. Tests have shown that the quality of the visualization may be related to the quality of the query result. In the context of similarity trees, the use of multiple descriptors contributed to a better-organized visualization.

Digital imaging plays an important role in many human activities, such as agriculture and forest management, earth sciences, urban planning, weather forecasting, medical imaging and so on. Processing, exploring and visualizing the inconceivable volumes of such images has turned out to be progressively troublesome. The Content-Based Image Retrieval (CBIR) remains an important issue that finds potential applications, given the place that retrieving digital images similar to a user-defined specification or pattern in huge databases now occupies in the day-to-day. CBIR systems use visual information like color, shape and texture to represent images in feature vectors. In general, there is an inconsistency in the evaluation of similarity between images according to human perception and the results computed by CBIR systems, which is called Semantic Gap. One way to improve CBIR systems is by the addition of techniques to visualize and interact with CBIR regarding similarity and diversity criteria, where the user can participate more actively in the process and steer the results according to its needs. In this paper we present the Visual-Interactive k-NDN Method (ViK): a novel approach to visualize and interact with Content-Based Image Retrieval systems. This paper aims at making use of Visual Data Mining techniques applied to queries in CBIR systems, improving the interpretability of the measure of diversity, applied using fractal analysis, as well as the relevance of results according to the prior knowledge of the user. Therefore, the user takes an active role in the content-based image retrieval, guiding its result and, consequently, reducing the Semantic Gap. Additionally, a better understanding of the diversity and similarity factors involved in the query is supported by visualization and interaction techniques.

This paper proposes a new approach to improve similarity queries with diversity, the Diversity and Visually-Interactive method (DiVI), which employs Visual Data Mining techniques in Content-Based Image Retrieval (CBIR) systems. DiVI empowers the user to understand how the measures of similarity and diversity affect their queries, as well as increases the relevance of CBIR results according to the user judgment. An overview of the image distribution in the database is shown to the user through multidimensional projection. The user interacts with the visual representation changing the projected space or the query parameters, according to his/her needs and previous knowledge. DiVI takes advantage of the users' activity to transparently reduce the semantic gap faced by CBIR systems. Empirical evaluation show that DiVI increases the precision for querying by content and also increases the applicability and acceptance of similarity with diversity in CBIR systems.

The use of computational resources can improve the diagnosis of medical diseases as a second opinion. Due to the large amount of data obtained daily, incremental techniques have been proposed to process medical data stream. In this paper we present an incremental decision tree classifier called StARMiner Tree (ST), which is based on Very Fast Decision Tree (VFDT) technique, to mine medical data. Different from VFDT, our proposed method ST does not depend on the number of reading samples to split a node. Because of it, ST is less conservative and describes the data since their first samples, being appropriate to be employed in medical environment, where not always a large number of data samples are available. We applied ST to four medical datasets, comparing the ST performance to the VFDT. The results indicated that ST is well-suited to deal with medical data streams, presenting high accuracy and low execution time.

Content-Based Image Retrieval (CBIR) Systems allow the search of images by similarity employing a numeric representation automatically or semi-automatically obtained from them to perform the search. Nevertheless, the query result does not always bring what the user expected. In this sense, CBIR systems face the semantic gap problem. One way of overcoming this problem is by the addition of diversity in query execution, so that the user can ask the system to return the most varied images regarding some similarity criteria. However, applying diversity on large datasets has a prohibitive computational cost and, moreover, the result often differs from the expected with a resulting subset that has images with high dissimilarity to the query image. In this paper we propose an approach to reduce the computational cost of Content-Based Image Retrieval systems regarding similarity and diversity criteria. The proposed approach employs dataset fractals analysis to estimate a suitable radius for a database subset to perform a similarity query regarding diversity. It selects closer images to the query center and applies the diversity factor to the subset, providing not only a better comprehension of the impact of the diversity factor to the query result, but also an improvement in execution time.

Data warehouse (DW) is an important component of Business Intelligence used to support strategic decision making. DW is a subject-oriented, nonvolatile, historical and massive database, which the processing of analytical queries, results in high response times. There known techniques for improving the performance processing of queries on DW. Among them is the use of data fragmentation, materialized views and indices. In addition, the NoSQL is an emerging technology whose main characteristics are improved query processing and data storage, and an alternative to relational databases. In this paper we investigate and compare the implementation of DW using relational databases and NoSQL, considering the Star Schema Benchmark. The results showed that the column-oriented model of the software FastBit showed a better performance, with gains of 25.4% to 99.4% if compared to other models NoSQL and the relational model, in the processing of queries on DW.

In this paper, we proposed a classifier based on statistical association rules that avoids the discretization step and automatically estimates the input thresholds. The algorithm automatically selects the most significant features to produce rules. These rules are simple, including the selected features, a single interval in the antecedent of the rule and a label class in the consequent, and getting at most twice the number of rules features. To evaluate our method, we compare it with traditional classifiers as C4.5 and Adaboost, in the task of classifying benign or malign masses of mammograms, usingtwo different real datasets. The proposed method achieve the best results regarding accuracy, sensitivity and sensibility.

Content-based image retrieval (CBIR) systems still face the problem of low precision of system results. To improve the precision of such systems, many image visual extractors have been developed and employed to represent the images. However, the usage of a large number of extractors and consequently, a large number of features, leads to the “dimensionality curse”, where the retrieval performance and the query accuracy diminish. In this paper, we propose a new method, called Statistical Fractal-scaled Product Metric (SFPM), to maximize the accuracy of CBIR systems and speedup similarity queries. The SFPM method combines association rule mining and the Fractal-scaled Product Metric (FPM) [4], to determine a reduced set of features and appropriate scale factors in multi-descriptor image similarity assessment. The FPM is an unsupervised method to determine a scale factor among features in multi-descriptor image similarity assessment based on the Fractal Theory. Experiments have shown that SFPM reduced the feature vector size in up to 65% and improved in up to 27% the query precision when comparing with the use of the FPM technique. The results show that the proposed method SFPM is effective in determining a reduced set of features and a near-optimal set of scale factors for the descriptors involved, and it is well-suited to improve the quality of content-based query in CBIR systems.

The ranking problem is a crucial task in the information retrieval systems. In this paper, we take advantage of single valued ranking evaluation functions in order to develop a new method of genetic feature selection tailored to improve the accuracy of content-based image retrieval systems. We propose to boost the feature selection ability of the genetic algorithms (GA) by employing an evaluation criteria (fitness function) that relies on order-based ranking evaluation functions. The evaluation criteria are provided by the GA and has been successfully employed as a measure to evaluate the efficacy of content-based image retrieval process, improving up to 22% the precision of the query answers. Experiments on three medical datasets containing breast cancer diagnosis and breast tissue density analysis showed that fitness functions based on ranking evaluation functions occupy an essential role on the algorithms' performance, obtaining results significatively better than other fitness function designs. The experiments also showed that the proposed method obtains results superior than feature selection based on the traditional decision-tree C4.5, naive bayes, support vector machine, 1-nearest neighbor and association rule mining.

Feature selection can significantly improve the precision of content-based queries in image databases by removing noisy features or by bursting the most relevant ones. Continuous feature selection techniques assign continuous weights to each feature according to their relevance. In this paper, we propose a supervised method for continuous feature selection. The proposed method applies statistical association rules to find patterns relating low-level image features to high-level knowledge about the images, and it uses the patterns mined to determine the weight of the features. The feature weighting through the statistical association rules reduces the semantic gap that exists between low-level features and the high-level user interpretation of images, improving the precision of the content-based queries. Moreover, the proposed method performs dimensionality reduction of image features avoiding the "dimensionality curse" problem. Experiments show that the proposed method improves the precision of the query results up to 38%, indicating that statistical association rules can be successfully employed to perform continuous feature selection in medical image databases.

In this paper we present a new method, called IDEA, which employs association rules to assist in medical image diagnosis. IDEA mines association rules, relating visual features with the knowledge gotten from specialists, and employs the associations to suggest possible diagnoses for a given medical image. IDEA incorporates two new algorithms called Omega and ACE. Omega performs simultaneously feature selection and data discretization very efficiently with linear cost on the number of feature values. ACE is a new associative classifier, which has the particular ability of suggesting multiple keywords to compose the diagnosis for a given medical image. The IDEA method has an important characteristic that makes it different from other CAD methods: it suggests multiple diagnosis hypotheses for an image and ranks them based on a measure of quality. The IDEA method was implemented in a prototype (IDEA system) for radiologists evaluate it. The radiologists showed enormous interest in employing the system to aid them in their daily work. The IDEA system was applied to real datasets and the results presented high accuracy (up to 96.7%). The results testify that association rules are well-suited to support the diagnosing task.

In this paper, we propose a method based on association rule-mining to enhance the diagnosis of medical images (mammograms). It combines low-level features automatically extracted from images and high-level knowledge from specialists to search for patterns. Our method analyzes medical images and automatically generates suggestions of diagnoses employing mining of association rules. The suggestions of diagnosis are used to accelerate the image analysis performed by specialists as well as to provide them an alternative to work on. The proposed method uses two new algorithms, PreSAGe and HiCARe. The PreSAGe algorithm combines, in a single step, feature selection and discretization, and reduces the mining complexity. Experiments performed on PreSAGe show that this algorithm is highly suitable to perform feature selection and discretization in medical images. HiCARe is a new associative classifier. The HiCARe algorithm has an important property that makes it unique: it assigns multiple keywords per image to suggest a diagnosis with high values of accuracy. Our method was applied to real datasets, and the results show high sensitivity (up to 95%) and accuracy (up to 92%), allowing us to claim that the use of association rules is a powerful means to assist in the diagnosing task.

In this paper we present HEAD, the Human Encephalon Automatic Delimiter, a new and efficient method for skull-stripping in T1-weighted MRI that combines an unique histogram analysis with binary mathematical morphology. In our experiments we use real images with highly variable noise ratios and intensity non-uniformity. We evaluate our results based on manually generated true masks and the well known Jaccard metric, achieving accuracy close to 99%. We compare our method with the popular Brain Extractor Surface algorithm (BSE), which in the same experiments achieved less than 95% of accuracy.

In this paper we present a framework based on association-rules to help diagnosis of mammogram abnormalities. Our framework - SuGAR - combines low-level features automatically extracted from images with high-level knowledge gotten from specialists to mine association rules, suggesting possible diagnoses. Our framework is optimized, in the sense that it combines, in a single step, feature selection and discretization, reducing the mining complexity. The framework was applied to real datasets and the results show high sensitivity (up to 95%) and accuracy (up to 92%), allowing us to claim that association rules can effectively aid in the diagnosing task.

This work aims at developing an efficient support to improve the precision of medical image retrieval by content, introducing an approach that combines techniques of statistical association rule mining and relevance feedback. Low level features of shape and texture are extracted from images. Statistical association rules are used to select the most relevant features to discriminate the images, reducing the size of the feature vectors and eliminating noisy features that influence negatively the query results, making the whole process more efficient. Additionally, our approach uses a new relevance feedback technique to overcome the semantic gap that exists between low level features and the high level user interpretation of images. Experiments show that the combination of statistical association rule mining and the relevance feedback technique proposed here improve the precision of the query results up to 100%.

This work aims at developing an efficient support for retrieval and classification of medical images, introducing an approach that comprises techniques of image processing, data mining and fractal theory, leading to an effective and direct way to compare images. A method of feature extraction and comparison is proposed, which uses Zernike moments for invariant pattern recognition as shape features of images' regions of interest. A new algorithm that generates statistical-based association rules is used to identify representative features that discriminate the disease classes of images. In order to minimize the computational effort, another new algorithm, based on fractal theory, is applied to reduce the dimensionality of the representative feature space. In essence, the proposed method determines the smallest set of relevant features that can properly represent images without loss of precision. In addition, the method discards the need of image segmentation, leading to a simple but effective way to make image retrieval by content. Experiments executing k-nearest neighbor queries on medical images reveal that the process is robust and suitable to perform retrieval combined with classification of this kind of images.