Texture Based Classification of Malaria Parasites from Giemsa-Stained Thin Blood Films

Abstract

Quantification of parasitemia is an important part of a microscopic malaria diagnosis. Giemsa-stained thin blood smear is the gold standard method for detecting malaria parasite enumeration. However, manual counting reveals the limitations of human inconsistency and fatigue, as well as the unreliability of accuracy and non-reproducibility. Inaccurate parasitemia affects clinical diagnosis and therapeutic procedure. Automated quantification is therefore useful to improve the performance of quantifying parasite density. In this paper, the texture-based classification approach is investigated. The methods consist of the following processes: pre-processing, segmentation, feature extraction and the classification of erythrocytes. The pre-processing is applied for image conversion and enhancement. The segmentation combines local adaptive thresholding, morphological process and watershed transform to extract red blood cells, separate touching and overlapping cells. Texture analysis is performed to establish parameters obtained from first-order, second-order and higher-order statistical analysis and wavelet transform. Two feature selection approaches, the sequential forward selection method and sequential backward elimination method, integrated with a support vector machine classifier are examined to obtain the optimal feature set for identifying the Plasmodium falciparum stages. We found that gray-level co-occurrence matrices based textural features were highly selected. The optimal feature set selected by the sequential forward selection yields lesser number of features and tends to give a higher degree of accuracy than the feature set selected by sequential backward elimination. The proposed method produces 98.87% accuracy for binary classification, 99.56% accuracy for ring stage classification, and 99.48% accuracy for tropozoite stage classification. Grey-level co-occurrence matrices based texture analysis is the dominant method compared to first-order and higher-order statistical texture analysis as well as wavelet transform.