Abstract:
Detecting brain tumors using medical images has improved a lot, but sharing data between different hospitals for research is not easy because of privacy concerns. Federated learning is a new way to solve this problem. In this thesis, we look at how well it works for brain tumor detection, with MLP model. We used a dataset with 44 different types of brain tumors and created a system with ten hospitals, each having a small computer (local device), and one central computer (global model). Each hospital trained its own model with its data and shared what it learned with the central computer. This way, the central computer learned from all the hospitals without seeing private patient data. The results showed that individual hospital models did very well, with the best accuracy reaching an impressive 98.5 percent! However, when we combined all the hospital models
in the central computer, the accuracy dropped to 77.2 percent. This happened because the data from each hospital was a bit different, making it hard for the central computer to put them all together perfectly. This thesis demonstrates that using federated learning for brain tumor detection is possible. However, it also highlights the need to find better ways to deal with the differences in hospital data. By doing this, we can improve brain tumor detection systems in the future while keeping patient data safe.
