Brain tumor is a fatal disease and one of the major causes of rising death rates in adults. Predicting methylation of the O6-Methylguanine-DNA Methyltransferase (MGMT) gene status utilizing Magnetic resonance imaging (MRI) imaging is highly important since it is a predictor of brain tumor responses to chemotherapy, which reduces the number of needed surgeries. Deep Learning (DL) approaches became powerful in extracting meaningful relationships and making accurate predictions. DLbased models require a large database and accessing or transferring patient data to train the model. Federated machine learning has recently gained popularity, as it offers practical solutions for data privacy, centralized computation, and high computing power. This study aims to investigate the feasibility of federated learning (FL) by developing a FL-based approach to predict MGMT promoter methylation status using the BraTs2021 dataset for the four sequence types, (Fluid Attenuated Inversion Recovery (FLAIR), T1-weighted (T1w), T1-weighted Gadolinium Post Contrast (T1wCE/T1Gd), and T2-weighted (T2w)) MRI images. The FL model compared to the DL-based and the experimental results show that even with imbalanced and heterogeneous datasets, the FL approach reached the training model to 99.99% of the model quality achieved with centralized data after 300 communication rounds between 10 institutions using OpenFl framework and the improved EfficentNet-B3 neural network architecture.