Developing pixel-based feature sets for intelligent identification of Eimeria species from microscopic images

Abdalla, Mohamed (2018) Developing pixel-based feature sets for intelligent identification of Eimeria species from microscopic images. Doctoral thesis, Northumbria University.

Preview

Text (Doctoral thesis) Abdalla.Mohamed_phd.pdf - Submitted Version Download (4MB) | Preview

Abstract

Morphological features have been investigated in various automation studies to identify different types of medical images. These studies rely on this type of feature, whereas digital images can provide other different features as descriptive characteristics. Pixels can be utilized as feature sets to recognize regions of interests. This research develops pixel values to extract informative features to identify protozoan parasites of the Eimeria genus. Eimeria is a single-celled intestinal parasite which infects humans and animals. Each type of host can be infected with different Eimeria species. Coccidiosis is caused when Eimeria infects animals, which is a rapidly spreading and fatal disease. Its treatment requires the identification of which species has infected the host, but similarities between Eimeria species make identification a very challenging process. Previously, automatic identification was carried out by imitating biological measurements, but these require complex and costly computational processes to extract the desired features. Therefore, this research aims to simplify the feature extraction process considering the use of another type of feature to distinguish between Eimeria species. The features considered do not need complex extraction processes and provide high accurate results. Pixel-based features are analysed by calculating the means of image matrix columns and rows of regions of interests. Features are represented as sets of column features (CF), row features (RF), and combinations of both in (CRF). Moreover, CF, RF, and CRF are extracted from greyscale level and colour images. Therefore, six feature sets are considered, and these are optimized by utilizing five selection and reduction algorithms to minimize the feature space. Furthermore, the extraction of super-pixel feature sets is developed to simplify segmentation. For classification, three classifiers are applied. The 5-fold cross-validation is used to evaluate the results and every experiment is repeated 50 times. Consequently, the results shown are the averages of 50 runs along with values of standard deviation. The proposed method is examined by analysing two microscopic image databases of 4402 images of the 7 Eimeria species in chickens and 2902 images of the 11 Eimeria species in rabbits. The best accuracy results achieved are 96.7% (±0.89%) and 95.85% (±2.4%) for the respective datasets. Finally, the proposed method succeeds in finding simple features to identify Eimeria species, reducing the feature number by 40% of the original size, and super-pixel feature sets are established which give excellent results.

Actions (login required)

View Item

Downloads