Blood cell classification using deep learning

• Main Author: Liaw, Mun Kin
• Format: Final Year Project / Dissertation / Thesis
• Published: 2022
• Subjects: T Technology (General)
• Online Access: http://eprints.utar.edu.my/4702/1/fyp_CS_2022_LMK.pdf; http://eprints.utar.edu.my/4702/

Throughout the years, hematology, which is the study of blood and its disorders, has played a big part in the health sector of Malaysia. Various diseases which could not be diagnosed physically could be overcame through analyzing blood samples of patients with the help of a microscope. However, the process was time-consuming and tedious as huge amounts of blood samples were needed to be manually checked by experts to achieve results which might not be clearly accurate due to human errors. The advancement of Artificial Intelligence (AI) has introduced complex methods such as deep learning that would automate the classification of blood cells in a fast and accurate manner Thus, the study of White Blood Cells (WBCs) classification using deep learning techniques is proposed in this research. The sole objective of the continuation of this project II is to define an efficient WBC classification model from scratch. The motivation was gotten from the literature review section where various researchers developed their own methods manually through experimenting such as ensemble methods, learning algorithms, combined methodologies, etc. Therefore, two ResNet models, ResNet-18 and ResNet-50 were built from scratch as the building blocks were studied and coded into the whole architecture which will run through several experiments of different configurations (batch size, learning rate, pooling layer, number of classes and number of dense layers). The learning rate was fine-tuned using the Tuner library from Keras to find the optimal value that generates a well-established model. As a result, the finalized model which consists of batch size of 16, learning rate of 0.0009, global pooling layer, 4 number of classes dropped to 3 number of classes and 3 dense layers achieved a testing accuracy of approximately 63% and validation accuracy of 99%.