ABSTRACT
With groundbreaking innovations in healthcare due to advancements in machine learning and deep learning, the need for computationally efficient and privacy-preserving solutions is exigent. Modern healthcare systems collect healthcare data in huge volumes. However, this data resides in different locations and is held by different institutions. This hinders its utilization by machine learning algorithms since they require large data sets from training. Moreover, accessing sufficient data from varied data silos is not possible due to various constraints like ethical, legal, economic, and technical challenges. Since medical data is highly sensitive, privacy concerns assume paramount importance. While dealing with similar scenarios, Federated Learning has shown promising results and could also be used in the healthcare domain for solving these issues. To explore the feasibility of federated learning in medicine by facilitating multi-institutional collaborations, the problem of brain tumor segmentation is undertaken. With the help of federated learning, training of models collaboratively from multiple parties owning the data while keeping privacy in check can be achieved. An aggregate ensemble of four different segmentation U-Net models for Cloud-based horizontal federated learning architecture with a FedProx algorithm is suggested to tackle the problem. In order to locate the area where the tumor is present, an average of codes is considered after every result to form the end MRI. This approach brings the model to the sensitive data owners and requires only the model updates to be sent to the central server, thereby preventing leakage of data and preserving privacy. While doing so, faulty updates from malicious clients can be discarded, thereby preventing them from hindering the central model. The challenges and considerations that need to be addressed while dealing with healthcare data for brain tumor segmentation are also highlighted in the chapter.
