CCE Theses and Dissertations

Date of Award

2020

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

College of Computing and Engineering

Advisor

Michael J. Laszlo

Committee Member

Francisco J. Mitropoulos

Committee Member

Sumitra Mukherjee

Keywords

circular Hough Transform (CHT), computer science, pulse coupled neural network (PCNN), red blood cells (RBCs), separation algorithm, white blood cells (WBCs)

Abstract

This research presents novel methods for segmenting digital blood cell images under a Pulse Coupled Neural Network (PCNN) framework. A blood cell image contains different types of blood cells found in the peripheral blood stream such as red blood cells (RBCs), white blood cells (WBCs), and platelets. WBCs can be classified into five normal types – neutrophil, monocyte, lymphocyte, eosinophil, and basophil – as well as abnormal types such as lymphoblasts and others. The focus of this research is on identifying and counting RBCs, normal types of WBCs, and lymphoblasts. The total number of RBCs and WBCs, along with classification of WBCs, has important medical significance which includes providing a physician with valuable information for diagnosis of diseases such as leukemia.

The approach comprises two phases – segmentation and cell separation – followed by classification of WBC types including detection of lymphoblasts. The first phase presents two methods based on PCNN and region growing to segment followed by a separate method that combines Circular Hough Transform (CHT) with a separation algorithm to find and separate each RBC and WBC object into separate images. The first method uses a standard PCNN to segment. The second method uses a region growing PCNN with a maximum region size to segment.

The second phase presents a WBC classification method based on PCNN. It uses a PCNN to capture the texture features of an image as a sequence of entropy values known as a texture vector. First, the parameters of the texture vector PCNN are defined. This is then used to produce texture vectors for the training images. Each cell type is represented by several texture vectors across its instances. Then, given a test image to be classified, the texture vector PCNN is used to capture its texture vector, which is compared to the texture vectors for classification.

This two-phase approach yields metrics based on the RBC and WBC counts, WBC classification, and identification of lymphoblasts. Both the standard and region growing PCNNs were successful in segmenting RBC and WBC objects, with better accuracy when using the standard PCNN. The separate method introduced with this research provided accurate WBC counts but less accurate RBC counts. The WBC subimages created with the separate method facilitated cell counting and WBC classification. Using a standard PCNN as a WBC classifier, introduced with this research, proved to be a successful classifier and lymphoblast detector. While RBC accuracy was low, WBC accuracy for total counts, WBC classification, and lymphoblast detection were overall above 96%.

Share

COinS