Abstract:
Acute Lymphoblastic Leukaemia is a fatal disease that affects white blood cells and bone marrow in
the human body. Every year, considerably a large number of adolescents and children become
victims of this type of leukaemia. The early detection of this disease directly affects the recovery rate
of the patients. In the manual process, pathologists can identify Acute Lymphoblastic Leukaemia and
the accuracy of the prediction may rely upon their experience. Hence this research has proposed an
image processing approach for early detection of Acute Lymphoblastic Leukaemia cells to prevent
the spreading of cancer, enabling the medical experts to initiate the treatment without any delay and
increase the recovery rate of such patients. For that, microscopic blood sample images were analyzed
considering the features such as color, shape, presence of nucleoli, and nucleon to a cytoplasmic ratio
of the cells separately using three Conventional Neural Networks (CNNs). Based on that, the Acute
Lymphoblastic Leukaemia cells were identified and classified as either Acute Lymphoblastic
Leukaemia or healthy. Compared to the laboratory testing methods, this approach obviously leads to
early detection of Acute Lymphoblastic Leukaemia with an accuracy of 94.57% that has been
confirmed by the domain experts. The proposed approach is an effective and less expensive method
that would assist doctors to get fast and accurate results. Hence the originality of this research was to
identify the presence of Acute Lymphoblastic Leukaemia cells in the microscopic blood sample
images and classify them as either Acute Lymphoblastic Leukaemia or healthy by identifying the
features of the Acute Lymphoblastic Leukaemia cells separately. Moreover, this research has found
that Conventional Neural Networks (CNN) is the most suitable Neural Network to identify Acute
Lymphoblastic Leukaemia using image processing technique.
Keywords: Acute Lymphoblastic Leukaemia; white blood cells; conventional neural networks; Image
Processing; Machine Learning
