Identification of Rabbit Coccidia by Using Microscopic Images

Abdalla, Mohamed A. E. and Seker, Huseyin and Jiang, Richard (2016) Identification of Rabbit Coccidia by Using Microscopic Images. In: 2016 International Conference on Engineering & MIS (ICEMIS) :. IEEE. ISBN 9781509055791

Abstract

Coccidia is an intestinal parasite that infects animals and causes Coccidiosis disease. Substantial animal mortality can be faced within several days of infection if it is not diagnosed or treated at the early stages of infection. Therefore, an urgent diagnostic tool has become necessary to tackle the spread of disease and to avoid animal death and subsequent business losses. Vets detect the disease by examining animal stool slides under microscope. They consider Eimeria oocyst sizes, textures and shapes to identify which genus has infected animals. However, this manual process is generally a challenging task as those oocysts are quite similar to each other and difficult to distinguish them. Apart from the morphological characteristics, this paper applies an automated, more robust and simple method to identify Coccidia species. It adopts pixel-based features instead of the morphological ones. Grey-scale images of Coccidia microscopic slides have been analyzed and the mean of their pixel values has been calculated to form three different feature sets to characterize the microscopic images. They are (i) column feature set (CF), (ii) row feature set (RF) and (iii) combination of both feature sets (CRF). Automated classification of the microscopic images is then carried out using K-Nearest Neighbor classifier. A five-fold cross validation is adapted to assess the robustness of the method and repeated for 50 times to avoid any statistical bias. As a case study, a database of 2902 microscopic images of Eimeria taken from rabbits has been analyzed, which yields a predictive accuracy of 70.44%, 67.47%and 80.13% for CF, RF and CRF, respectively. Furthermore, the number of features has been reduced by eliminating low degree of the features. This process has reduced the feature size by as much as 30% and improved the best predictive accuracy by about 3% to 82.83%. This promising outcome is expected to lead a fully automated mobile diagnostic tool for parasite detection.