Automatic detection of pathological retinal images using color and shape features

Flávio Henrique Duarte de Araújo, Rodrigo de Melo Souza Veras, Romuere Rodrigues Veloso e Silva, André Macedo Santana, Fátima Nelsizeuma Sombra de Medeiros

Resumo


Objective: We propose an algorithm for exudate detection and pathological retinal images identification. Method: We improved an existing algorithm that detects exudates in a retinal image replacing the k-means clustering by fuzzy k-means and applied an additional step to detect optical disc (OD). Furthermore, our approach added a classification process to eliminate remaining false exudate regions. Finally, we classify the retinal image as pathological or non-pathological by measuring the ratio of candidate exudate regions before classification and the number of regions removed by the classification step. Results: Tests were performed on DIARETDB1 database, and the results obtained were; Fmeasure – 90%, area under the ROC curve – 88% and the Kappa coefficient – 77% (very good). Conclusion: The success of the algorithm is due mostly to the OD detection approach and the classification step. The obtained results confirmed that the proposed algorithm outperformed the others.


Palavras-chave


Machine Learning; Diagnostic Imaging; Exudates and Transudates

Texto completo: PDF


Journal of Health Informatics - ISSN 2175-4411
Rua Tenente Gomes Ribeiro, 57 - sala 33 CEP 04038-040 São Paulo - SP - Brasil
Tel./Fax: + 55 11 3791 3343 - E-mail: jhi@sbis.org.br