Tools
Tools
Oduntan, Ifeoma (2022) Analysis of Microscopic Blood Images in Sickle Cell Classification Using Deep Learning Algorithm. Masters thesis, Dublin, National College of Ireland.
Preview |
PDF (Master of Science)
Download (2MB) | Preview |
Preview |
PDF (Configuration manual)
Download (3MB) | Preview |
Abstract
Abstract. Sickle cell disease is the most common form of hereditary blood disorder that is associated with hemoglobin abnormality due to mutation in β-globin genes known as hemoglobin S. It is estimated that 20 million people around the world live with the disease and a total of 176,000 deaths were recorded in 2013. The traditional method of diagnosing it is through conventional analysis of peripheral blood smears under the microscope by a pathologist which is laborious, time consuming and can lead to delays and misdiagnosis. Currently, the conventional machine learning technique still depends on the expertise knowledge of medical practitioners to select the features, and this can affect the classifier’s accuracy due to the subjective nature of the process. Developing an automatic way of diagnosing this disease through classification of the red blood cells as early as possible is a challenge due to lack of data in the medical field. This research aims to apply a deep learning technique that implements a novel Deep Convolutional Generative Adversarial Networks (DCGANs) for image synthesis to overcome small dataset issue for efficient classification and diagnosis of sickle cell disease. The augmented erythrocytesIDB1 dataset is used as an input to DCGANs to generate more images which can be used to train six deep transfer learning image classification models namely DenseNet121, ResNet50, InceptionV3, VGG16, VGG19, and MobileNet based on three types of red blood cells namely circular (normal), elongated (sickle cells), and other abnormality. The performance of the models is compared on the original images, GAN generated images/original images, and the traditional augmented images/original images to see the effect of each dataset on each model and find out if GAN generated images are realistic and can be an alternative source for augmenting data for classification in situation where the data size is very small, especially in the medical field and also identify the optimal classification model. The results are presented based on weighted metrics of accuracy, precision, recall, and F1-score and it showed that model performance on GAN generated images improved between 4.5% to 136% in all the models and MobileNet model achieved the highest accuracy and recall of 99.70%.
| Item Type: | Thesis (Masters) |
|---|---|
| Uncontrolled Keywords: | Microscopic blood images; Generative Adversarial Networks; Data Augmentation; Red blood cells; Sickle Cell Disease; Transfer learning; Deep learning |
| Subjects: | Q Science > QA Mathematics > Electronic computers. Computer science T Technology > T Technology (General) > Information Technology > Electronic computers. Computer science R Medicine > R Medicine (General) R Medicine > RB Pathology Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Machine learning |
| Divisions: | School of Computing > Master of Science in Data Analytics |
| Depositing User: | Tamara Malone |
| Date Deposited: | 27 Feb 2023 15:32 |
| Last Modified: | 02 Mar 2023 08:27 |
| URI: | https://norma.ncirl.ie/id/eprint/6243 |
Actions (login required)
 |
View Item |