Blood Cancer Diagnostic System using Deep Learning

¹Ume-Ezeoke Jenevieve Chika
² Nweze, Rosemary Chika
³Nkanu, Inyang Unoh
⁴Ogah, Ogbonnaya Sunday
⁵Ugwuanyi Sunday Chinedu

¹Computer Science Department, Ebonyi State University Abakaliki,
^2,5Computer Science Department, State University of Medical & Applied Sciences, Igbo-Eno, Enugu State, Nigeria.
³ICT Unit, Ministry of Agriculture, Calabar, Cross River State, Nigeria.
⁴Department of Computer Science & Mathematics, Evangel University, Ebonyi State, Nigeria.

Abstract

Blood cancer is currently the seventh most common type of cancer and the ninth leading cause of cancer-related deaths in both males and females. In medical diagnosis, time and accuracy are critical factors. Traditional pathological diagnostic methods are often expensive, time-consuming, and prone to human error. As a result, the use of deep learning algorithms for detection and diagnosis is becoming increasingly prevalent due to their high performance and accuracy. We developed a system that effectively diagnoses two types of blood cancer, leukemia and lymphoma. In order to test the accuracy of lymphoma and leukemia diagnosis, other kinds of cells were added into our dataset such as lymphocytes, which are not very easy for doctors to distinguish. Each type of blood cancer has distinct characteristics. Deep learning techniques enable the development of automated and effective systems for their identification. In this study, pre-trained convolutional neural networks (CNNs) including ResNet50 (a 50-layer Residual Network), Google Inception, and VGG16 (a 16-layer model from the Visual Geometry Group) were utilized through transfer learning. These models were fine-tuned using a dataset of blood cell images and trained using an end-to-end method. An ensemble model was developed by averaging the predictions of the three individual classifiers. All models were evaluated, and a comparative analysis was conducted to determine the most effective classifier. This ensemble approach achieved higher accuracy of 91%, sensitivity 100%, f1 score 100% and precision 95% which outperformed the individual models. The final diagnostic system was deployed as an offline application to ensure accessibility for medical personnel.

Keywords: Blood cancer, Leukemia, Lymphoma, Lymphocytes, Medical diagnosis, Cancer detection, Diagnostic accuracy, Sensitivity, Precision, F1 score

Technical & Methodological Keywords: Deep learning,  Convolutional neural networks (CNNs), Transfer learning, ResNet50, Google Inception, VGG16, Ensemble model, End-to-end training, Image classification, Model evaluation

Blood-Cancer-Diagnostic-System-using-Deep-Learning