Flávio H. D. Araújo

Leukemia diagnosis in blood slides using transfer learning in CNNs and SVM for classification

Abstract Leukemia is a pathology that affects young people and adults, causing premature death and several other symptoms. Computer-aided systems can be used to reduce the possibility of prescribing inappropriate treatments and assist specialists in the diagnosis of this disease. There is a growing use of Convolutional Neural Networks (CNNs) in the classification and diagnosis of medical image problems. However, the training of CNNs requires a large set of images. To overcome this problem, we use transfer learning to extract images features for further classification. We tested three state-of-the-art CNN architectures and the features were selected according to their gain ratios and used as input to the Support Vector Machine classifier. The proposed methodology aims to correctly classify images with different characteristics derived from different image databases and does not require a segmentation process. We built a new database from the union of three distinct databases presented in the literature to validate the proposed methodology. The proposed methodology achieved hit rates above 99% and outperformed nine methods found in the literature.

A multi-objective approach for calibration and detection of cervical cells nuclei

The automation process of Pap smear analysis holds the potential to address womens health care in the face of an increasing population and respective collected data. A fundamental step for automating analysis is cell detection from light microscopy images. Such information serves as input to cell classification algorithms and diagnostic recommendation tools. This paper describes an approach to nuclei cell segmentation, which critically impacts the following steps for cell analyses. We developed an algorithm combining clustering and genetic algorithms to detect image regions with high diagnostic value. A major problem when performing the segmentation of images is the cellular overlay. We introduce a new nuclear targeting approach using heuristics associated with a multi-objective genetic algorithm. Our experiments show results using a public 45-image dataset, including comparison to other cell detection approaches. The findings suggest an improvement in the nuclei segmentation and promise to support more sophisticated schemes for data quality control.

Unsupervised Leukemia Cells Segmentation Based on Multi-space Color Channels

Leukemia is a type of cancer that originates in the bone marrow and is characterized by abnormal proliferation of white blood cells. In order to have correct identification of lymphoblasts, hematologists examine blood blades of the patient. A low cost and efficient solution to facilitate the work of these experts is the use of systems to examine blood microscopic images. Segmentation is considered a crucial step to developing these systems. In this paper, we propose an automatic segmentation technique that uses two-color systems and the clustering algorithm K-means. The proposed approach is evaluated on three public image databases with different characteristics and performance measures used are: accuracy, specificity, sensitivity and Kappa index. The results obtained in the experiments have Kappa index of 0.9306 in ALL-IDB 2, 0.8603 in BloodSeg and 0.9119 in Leukocytes database. These measures outperform other methods of literature.

Optic disc detection in retinal images using algorithms committee with weighted voting

This paper presents a new method for Optic Disc (OD) detection in color retinal images. Processing and analyzing these images constitute a relevant task to help specialists in eye diseases detection. Particularly, finding OD in a retinal fundus image, improves significantly the chances to detect diseases. OD location serves as input to the detection of other retinal anatomical structures such as macula, blood vessels and some anomalies, such as exudates, hemorrhages and drusen. These anomalies will serve to determine the presence of retinal diseases. We have implemented five OD detection methods from state of art and created a committee of algorithms. Unlike other proposals, based on simple majority vote, the output of the proposed committee is established using a weighted voting obtained by each algorithm. For the definition of the weights we use a portion of available image databases and calculate the success rate of each of the five methods. Tests were carried on six public benchmark databases, which constitute a total of 1566 images.

ABCD rule and pre-trained CNNs for melanoma diagnosis

Skin cancer is the most common type of cancer and represents more than half of cancer diagnoses. Melanoma is the least frequent among skin cancers, but it is the most serious, with high potential for metastasis and can lead to death. However, melanoma is almost always curable if discovered in the early stages. In this context, computational methods for processing and analysis of skin lesion images have been studied and developed. This work proposes a computational approach to assist dermatologists in the diagnosis of skin lesions in melanoma or non-melanoma by means of dermoscopic images. The proposed methodology classifies skin lesions using a descriptor formed by the combination of the ABCD rule (Asymmetry, Border, Color, and Diameter) and pre-trained Convolutional Neural Networks (CNNs) features. The features were selected according to their gain ratios and used as input to the MultiLayer Perceptron classifier. We built a new database joining two distinct databases presented in the literature to validate the proposed methodology. The proposed method achieved an accuracy rate of 94.9% and Kappa index of 89.2%, which is considered “excellent”.

Reverse image search for scientific data within and beyond the visible spectrum

Abstract The explosion in the rate, quality and diversity of image acquisition instruments has propelled the development of expert systems to organize and query image collections more efficiently. Recommendation systems that handle scientific images are rare, particularly if records lack metadata. This paper introduces new strategies to enable fast searches and image ranking from large pictorial datasets with or without labels. The main contribution is the development of pyCBIR , a deep neural network software to search scientific images by content. This tool exploits convolutional layers with locality sensitivity hashing for querying images across domains through a user-friendly interface. Our results report image searches over databases ranging from thousands to millions of samples. We test pyCBIR search capabilities using three convNets against four scientific datasets, including samples from cell microscopy, microtomography, atomic diffraction patterns, and materials photographs to demonstrate 95% accurate recommendations in most cases. Furthermore, all scientific data collections are released.

Diagnosing Leukemia in Blood Smear Images Using an Ensemble of Classifiers and Pre-Trained Convolutional Neural Networks

Leukemia is a worldwide disease. In this paper we demonstrate that it is possible to build an automated, efficient and rapid leukemia diagnosis system. We demonstrate that it is possible to improve the precision of current techniques from the literature using the description power of well-known Convolutional Neural Networks (CNNs). We extract features from a blood smear image using pre-trained CNNs in order to obtain an unique image description. Many feature selection techniques were evaluated and we chose PCA to select the features that are in the final descriptor. To classify the images on healthy and pathological we created an ensemble of classifiers with three individual classification algorithms (Support Vector Machine, Multilayer Perceptron and Random Forest). In the tests we obtained an accuracy rate of 100%. Besides the high accuracy rate, the tests showed that our approach requires less processing time than the methods analyzed in this paper, considering the fact that our approach does not use segmentation to obtain specific cell regions from the blood smear image.

SURF Descriptor and Pattern Recognition Techniques in Automatic Identification of Pathological Retinas

Automatic systems for eye disease identification are important in the ophthalmology field. Medical image processing and analysis have enabled prevention, diagnosis and treatment of eye diseases which were assumed to incurable, in the past. Despite the advances in technology, automated systems for retina component detection are affected by the large diversity of images and degradation caused by artifacts originated by diseases. These diseases can potentially degrade images and cause inaccurate diagnosis for automated systems as well as specialists. Therefore, several researches have been employed to overcome these drawbacks. In this study, we introduce a computer-aided diagnosis system that can identify patients who present risks of vision loss and require greater assistance of a specialist. Our approach applies the Speed-Up Robust Feature algorithm to find points of interest to form visual dictionaries. These dictionaries describe an image as a vector of characteristics which is input to the step that classifies retina image as healthy or pathological. We evaluated the proposed system in 1674 images of six public databases and obtained 96.70% accuracy and Kappa index of 0.9 (excellent).

Ferramenta Computacional para o Ensino de Algoritmos de Ordenação

Este artigo discute a colaboracao entre estudantes na internet e a troca de conhecimento por meio desse canal de comunicacao. Propoe uma metodologia que visa potencializar as relacoes entre os usuarios de uma rede educacional situada num contexto especifico. Tal metodo tem como fundamento as teorias do contexto e a analise de redes sociais (ARS), para promover um mapeamento das redes estudadas.Refletir sobre curriculo escolar formal e comunidades de aprendizagem como metafora das TIC dinamizam os caminhos empiricos, construcoes criticas e aprofundamento da difusao do conhecimento como parte do processo de humanizacao/ tecnologizacao do homem. Oriundos de processos e movimentos contemporâneos, a consolidacao de ambientes computacionais nas escolas potencializa a construcao do conhecimento e a socializacao de praticas pedagogicas inovadoras. Esta investigacao assume as situacoes especificas curriculares, procura descobrir o que existe de mais essencial e caracteristico, partindo do conhecimento de curriculo e suas bases teoricas tradicionais, para a construcao da discussao sobre um curriculo em rede associada a instrumentalizacao das comunidades de aprendizagem.O sistema Moodle constitui-se atualmente numa das mais importantes ferramentas de apoio a cursos na Web. Apesar disto, seu modelo apresenta algumas deficiencias para uma estruturacao hierarquica e compartilhamento de materiais digitais entre disciplinas e turmas do seu ambiente virtual. Este artigo apresenta um estudo sobre a arquitetura central do Moodle, propondo a definicao de um novo nucleo, visando o aprimoramento destas caracteristicas.Estudo descritivo, qualitativo, com estudantes da 3a serie de Graduacao em Enfermagem de uma Universidade Publica de Sao Paulo, SP. Os participantes construiram Mapas Conceituais, por meio do software Cmap Tools®. Os dados foram coletados em um Grupo Focal e todos os sujeitos indicaram que o uso do software facilita e garante a organizacao, visualizacao e correlacao dos dados, porem houve dificuldades iniciais relacionadas ao manejo das ferramentas. Conclui-se, que o software Cmap Tools® favoreceu a construcao dos MC por seus recursos de formatacao, porem estrategias de orientacao deveriam ser implantadas. Como resultado, desenvolveu-se um manual para o uso do software Cmap Tool® em video Podcasting.Ha poucas iniciativas com respeito aos ambientes de virtuais para a divulgacao de materiais curriculares sobre modelagem matematica. Esses ambientes oferecem acesso as praticas pedagogicas em modelagem. Este trabalho apresenta um sistema Web para hospedar atividades de modelagem e materiais multimidia para descrever o desenvolvimento do ambiente de modelagem em sala de aula e apoiar outros professores na implementacao em suas praticas pedagogicas.A composicao e sequenciamento de Objetos de Aprendizagem sao discutidas neste trabalho a partir da representacao da estrutura conceitual de um dominio em termos das suas relacoes de dependencia. A composicao de Objetos de Aprendizagem e modelada a partir da estrutura narrativa de um discurso considerando-se os aspetos formais dos planos do conteudo e de expressao. O aspecto formal do conteudo da composicao e dado pelas pelos conceitos e seus relacionamentos e forma da expressao corresponde aos tipos de signos definidos pelo LOM. A estrutura da composicao obtida independe do tipo de midia utilizado e o modelo adequa-se as propostas de composicao adaptativas tanto do ponto de vista do meio como das disponibilidades de conexao.