Authors - Shiva Kumar Bandaru, Upendra Pratap Singh Abstract - In a federated learning-based setup, parameter aggregation plays a pivotal role in obtaining global parameter estimates that assimilate the knowledge learned by the different clients. With an efficient parameter aggregation strategy, the global parameter estimates derived are more generalizable, accelerating the local client training in the subsequent communication rounds. In the proposed approach, we propose a novel m-ary improvisation-based parameter aggregation algorithm to obtain the global parameters. Specifically, after a threshold number of communication rounds has elapsed, the performance of the clients is evaluated on an independent test set, and the clients with better generalization are labeled as strong and do not participate in the next set of a threshold number of communication rounds. In this way, weak clients participate in the federated learning for more communication rounds; after the next set of threshold communication rounds has elapsed, the clients undergo a similar evaluation to be labeled as strong or weak again. The proposed algorithm ensures weak clients get more attention/exposure to learn the model parameters collaboratively. The global model trained on the BraTS2020 dataset in a federated learning-based framework reports the Dice coefficient, Jaccard index and pixel accuracy values of 0.8851, 0.8965, and 99.92%, respectively. Further, we show empirically that the training time for the different clients reduces from 180 minutes in the first phase of federated learning to only 64.8 minutes in the last phase, highlighting an accelerated training process. Consequently, the results reported by the proposed federated learning-based segmentation model highlight its usability for efficiently carrying out brain segmentation involving private and sensitive brain scans.
Open Zoom
