CLASSIFICATION WITH MACHINE LEARNING METHODS FROM MULTI-SEQUENCE MR IMAGES OF CHILDHOOD POSTERIOR FOSSA TUMORS

Childhood brain tumors rank high among the leading causes of mortality, being the second most common type of cancer after leukemia. Abnormal structures in the brain are visualized using MRI techniques, which are the most commonly employed tools for distinguishing the neural structure of the human brain. However, identifying and diagnosing these abnormal structures can be a time-consuming and critical process. In this study, tumors in the Magnetic Resonance images of patients with Posterior Fossa tumors were segmented using two different image segmentation methods. Subsequently, numerical features were extracted from these tumors, and significant numerical features among tumor groups were determined using the Student's T-test; based on these features, tumor types were classified using machine learning algorithms. The study focused on the three most common types of Posterior Fossa tumors: Medulloblastoma, Ependymoma, and Pilocytic Astrocytoma, utilizing T2, Contrast-Enhanced T1, and ADC sequences. A total of forty-eight different numerical features were extracted from the segmented tumors and then acquired significant features were classified using five different machine learning algorithms. Among PA-MB, EM-MB and EM-PA tumor types, the average result of the most successful method in the T1 sequence was 86.93%, while it was 93.7% for the T2 sequence and 92.06% for the ADC sequence. Decision tree, SVM and Ensemble classifiers gave more successful results than others. As a result of the detailed examination, our study not only makes valuable contributions to the literature, but also has a promising structure in terms of its potential to help clinicians.