| dc.description.abstract | Segmentation and Classification of White Blood Cells for Automated Differential Count
Abstract and Synopsis
Introduction
Medical image analysis plays a major role in providing quality healthcare. Improved imaging techniques have enabled effective diagnosis of various diseases. Automation of diagnostic processes is essential because manual methods require considerable time, effort, and care, and are prone to human error. This thesis focuses on segmentation and classification of white blood cells (WBCs) for automation of the differential count (DC) of blood.
The DC involves classification and counting of five classes of WBCs in peripheral blood:
Lymphocytes
Monocytes
Eosinophils
Basophils
Neutrophils
While semi-automatic systems exist, this work aims to develop a fully automatic and reliable system for DC.
Methodology
Segmentation
Input: Color images of stained blood smears.
WBCs are extracted from the background (red blood cells, plasma, platelets, fragments).
Both nucleus and cytoplasm are segmented.
Proposed algorithm avoids parameter tuning and assumptions of circular shape or fixed color range.
Steps:
Convert RGB image to HSV space.
Perform K-means clustering to locate nuclei.
Subdivide into single-WBC images.
Apply clustering again, compute mean/variance parameters.
Initialize Expectation-Maximization (EM) algorithm for iterative segmentation.
Identify nucleus (high saturation) and cytoplasm (spatial proximity).
Feature Extraction
Shape-based features: nucleus eccentricity, cytoplasm eccentricity, nucleus compactness, cytoplasm-to-nucleus area ratio, number of nucleus lobes.
Color features: average red, green, blue components for nucleus and cytoplasm.
Textural features (cytoplasm only): energy, entropy, correlation, coarseness, busyness.
Classification
Classifiers used: Nearest Neighbour (NN), k-NN, Weighted k-NN, Bayes classifier, Neural Networks, Support Vector Machines (SVM).
Training set: 50 patterns (~10 representatives per WBC class).
Results: Neural Networks achieved ~97% accuracy, SVM ~94%, Bayes ~82%.
Conclusions
Developed a fully automatic system for WBC segmentation and classification.
Algorithm eliminates manual parameter tuning.
Neural Networks provided the highest classification accuracy.
System shows promise for reliable automation of blood differential count. | |
