Early Detection of ALL Disease Using YOLOv4 Algorithm on Peripheral Blood Smear Images

Abstract

Peripheral smear is a method that allows the analysis of blood with the help of a microscope after staining the blood liquid on a glass slide. The morphological appearance and multiplicity of some blood elements are examined by this method. This examination is preferred for evaluating the response to treatment as well as providing information about the diagnosis and severity of diseases. Since manual examination of blood elements by operators contains many humanistic and environmental parameters, it is time-consuming and error-prone. Instead of this approach which does not offer a standard accuracy rate, it is planned to create a quick, cost-effective and fully automatic computer-aided system. In this study, the examination of ALL(Acute Lymphoblastic Leukaemia) malignancy which is one of the main types of leukaemia by peripheral smear method will be handled. ALL is formed by the overproduction of malignant and immature white blood cells called lymphoblasts. Since it spreads to the blood circulation and vital organs rapidly, the speed of the tests for early diagnosis is important. The peripheral blood smear test, which is a part of the diagnosis process of ALL malignancy, is a guide in terms of diagnosis and severity of the disease with the presence and multiplicity of lymphoblast cells. Therefore, a fully automatic system is designed to determine the presence and number of blasts on the given image. The model used to create this system is trained with the YOLOV4 algorithm. Training is provided via peripheral blood smear images the free and public in the ALL-IDB dataset (https://homes.di.unimi.it/scotti/all/) tagged by expert oncologists. As a result of the test process, the success rate, precision, recall, specificity and F score criteria are obtained as 98.87%, 99.08%, 99.81%, 88.88% and 99.44%, respectively. With this work, it is real-time given the total number of blast cells on the image, which is an important biomarker in determining the severity of ALL disease or evaluating the response to treatment. It is expected that this fully automatic system which produces output after 3–4 seconds for the given input will be a helpful idea for the operators.