Lung disease detection using feature extraction and extreme learning machine

INTRODUCTION: The World Health Organization estimates that by 2030 the Chronic Obstructive Pulmonary Disease (COPD) will be the third leading cause of death worldwide. Computerized Tomography (CT) images of lungs comprise a number of structures that are relevant for pulmonary disease diagnosis and analysis. METHODS: In this paper, we employ the Adaptive Crisp Active Contour Models (ACACM) for lung structure segmentation. And we propose a novel method for lung disease detection based on feature extraction of ACACM segmented images within the cooccurrence statistics framework. The spatial interdependence matrix (SIM) synthesizes the structural information of lung image structures in terms of three attributes. Finally, we perform a classification experiment on this set of attributes to discriminate two types of lung diseases and health lungs. We evaluate the discrimination ability of the proposed lung image descriptors using an extreme learning machine neural network (ELMNN) comprising 4-10 neurons in the hidden layer and 3 neurons in the output layer to map each pulmonary condition. This network was trained and validated by applying a holdout procedure. RESULTS: The experimental results achieved 96% accuracy demonstrating the effectiveness of the proposed method on identifying normal lungs and diseases as COPD and fibrosis. CONCLUSION: Our results lead to conclude that the method is suitable to integrate clinical decision support systems for pulmonary screening and diagnosis.

Detecção de exsudatos em imagens de retina por técnicas de morfologia matemática e agrupamento nebuloso / Exudate detection in retina images by mathematical morphology techniques and fuzzy clustering

A retinopatia diabética (RD) é uma das principais complicações do diabetes mellitus, pois causa sérios danos à retina e consequentemente à visão, podendo inclusive resultar em cegueira. O diagnóstico da RD é realizado através da análise visual de imagens de retina, sendo os exsudatos (depósitos de gordura) os principais padrões rastreados pelo médico especialista. Vale destacar que o diagnóstico precoce, realizado através do monitoramento regular, associado ao tratamento adequado apresenta inúmeros benefícios na prevenção da deficiência visual. Neste trabalho, é proposto um algoritmo de detecção de exsudatos em imagens de retina, cuja validação experimental é realizada na base pública DIARETDB1. A escolha desta base se deve à disponibilidade da localização dos exsudatos na retina, o que constitui o padrão ouro para a validação dos algoritmos. A metodologia proposta combina agrupamento nebuloso e técnicas de morfologia matemática, além de prover a detecção do disco óptico considerando que o mesmo é um ponto de convergência dos vasos. Os resultados mostraram que o método de detecção de exsudatos apresentou taxas de acerto na avaliação por imagens e por regiões na ordem de 73,03% e 99,41%, respectivamente. Estes resultados confirmam que houve uma melhoria no desempenho na detecção, quando comparados, aos resultados de métodos disponíveis na literatura.

Diabetic retinopathy (DR) is one of the major complications of diabetes mellitus and, furthermore it causes severe damage to the retina and consequently to the vision. DR may lead to blindness and therefore it is important to prevent it or early detect and treat it. The diagnosis of DR is performed by visual analysis of retinal images being exudates (fat deposits) the main patterns traced by a specialist doctor. It is noteworthy that early diagnosis, through regular monitoring when coupled with proper treatment, results in numerous benefits in the prevention of visual impairment. Thus, this paper proposes an algorithm for exudate detection in retinal images, whose experimental validation is performed on retina images of the publicly available DIARETDB1 database. The reason for choosing this database is that it provides spatial coordinates of exudates in retina images which constitute ground truths for the algorithm validation. The proposed methodology combines fuzzy clustering and mathematical morphology techniques, and thus it provides a method for optic disk detection considering that it is as the convergent point of vessels. The exudate detection method presented successful rates of 73.03% and 99.41% concerning the use of the whole image and only partial regions, respectively. These results confirm the performance improvement provided by the proposed methodology, when comparing it to other methods available in the literature.