نازیلا کریمیان

Thesis: An automated method for diagnosis of thalassemia minor from blood slide microscopic images and red blood cell feature extraction

Abstract

Thalassemia is one of the most common genetic diseases all over the world. The disease is classified into two types of thalassemia minor and major. Thalassemia minor is not considered as a disease and lots of people who have got the disease are not aware of it. If two thalassemia minor patients get married, their child would be born with thalassemia major with a probability of 25%. Symptoms of thalassemia major gradually become evident in the first 6 months of baby’s life. Thalassemia major usually leads to death and patients with thalassemia major usually have a short lifetime. The cost of taking care of these patients is very high. To avoid the cost, premarital testing is required. In Iran since 1376, national thalassemia prevention program has been considerably established in health centers. Due to the high volume of these tsets, there is a need for automatic thalassemia detection approaches. Image processing techniques can help experts in this regard. Input data are microscopic images of blood smear slides which are taken by a digital camera connected to the microscope. In our research Blood slides were collected from Seyed-Alshohada Hospital (Esfahan, Iran) and Department of Medical Science of university of Esfahan. Images of 100 normal and 100 thalassemic blood samples have been prepared. The goal of this research is to design and implement a practical system for detecting thalassemia minor samples automatically from normal samples based on the morphology of RBCs. In most blood diseases, shape of RBCs change from normal (disk-like shape) to different shapes. Target cells, Elliptocytes and Dacrocytes (tear drop) are types of RBCs which exist in thalassemic blood samples. In this research these abnormal cells are identified and thalasssemia minor is detected based on the percentage of these RBCs. After data collection, in first step RBCs are segmented based on Otsu threshholding method. Then a set of appropriate geometric feartures are extracted from each RBC. RBCs are classified in three steps. In the first step, Target cells are detected by threshholding and accuracy of 84%. In the second step, Normal RBCs, Elliptocytes and Dacrocytes are classified with Back Propagation Neural Network, Support Vector Machine and K-Nearest Neighbor algorithm. K-nearest Neighbor algorithm had an acceptable performance and could achieve accuracy of 99% for normal RBCs, 99% for Elliptocytes and 94% for Dacrocytes. A threshold level has been selected for the classifier’s output and if the output is less than this threshold level, the RBC is eliminated. In the third step, these eliminated RBCs are considered as Poikilocytes. Detection of these cells is achieved with accuracy of 95%. After classificiation of RBC types, thalassemia minor is detected by using the percentage of classified RBCs (which are given to a neural network as input feature vector). It is shown that the ultimate algorithm used in this thesis could achieve detection of thalassemia minor with accuracy of 98% and sensitivity equal to 100%.

Keywords:

Thalassemia, Image Processing, Red Blood Cell morphology, Anemia detection