Abstract

In this work we develop a system for automatic detection and classification of cytological images which plays an increasing important role in medical diagnosis. A primary aim of this work is the accurate segmentation of cytological images of blood smears and subsequent feature extraction, along with studying related classification problems such as the identification and counting of peripheral blood smear particles, and classification of white blood cell into types five. Our proposed approach benefits from powerful image processing techniques to perform complete blood count (CBC) without human intervention. The general framework in this blood smear analysis research is as follows. Firstly, a digital blood smear image is de-noised using optimized Bayesian non-local means filter to design a dependable cell counting system that may be used under different image capture conditions. Then an edge preservation technique with Kuwahara filter is used to recover degraded and blurred white blood cell boundaries in blood smear images while reducing the residual negative effect of noise in images. After denoising and edge enhancement, the next step is binarization using combination of Otsu and Niblack to separate the cells and stained background. Cells separation and counting is achieved by granulometry, advanced active contours without edges, and morphological operators with watershed algorithm. Following this is the recognition of different types of white blood cells (WBCs), and also red blood cells (RBCs) segmentation. Using three main types of features: shape, intensity, and texture invariant features in combination with a variety of classifiers is next step. The following features are used in this work: intensity histogram features, invariant moments, the relative area, co-occurrence and run-length matrices, dual tree complex wavelet transform features, Haralick and Tamura features. Next, different statistical approaches involving correlation, distribution and redundancy are used to measure of the dependency between a set of features and to select feature variables on the white blood cell classification. A global sensitivity analysis with random sampling-high dimensional model representation (RS-HDMR) which can deal with independent and dependent input feature variables is used to assess dominate discriminatory power and the reliability of feature which leads to an efficient feature selection. These feature selection results are compared in experiments with branch and bound method and with sequential forward selection (SFS), respectively. This work examines support vector machine (SVM) and Convolutional Neural Networks (LeNet5) in connection with white blood cell classification. Finally, white blood cell classification system is validated in experiments conducted on cytological images of normal poor quality blood smears. These experimental results are also assessed with ground truth manually obtained from medical experts.