PROJECT TITLE :

SplitAVG A heterogeneity-aware federated deep learning method for medical imaging

ABSTRACT:

An emerging research paradigm known as federated learning enables the training of deep learning models in a collaborative setting without requiring the sharing of patient data. However, the data from various institutions are typically different from one another, and this can cause the performance of models that have been trained using federated learning to suffer. In this investigation, we propose a novel heterogeneity-aware federated learning method called SplitAVG as a means of overcoming the performance drops that can result from federated learning due to data heterogeneity. In contrast to earlier federated methods, which necessitated intricate heuristic training or hyper parameter tuning, our SplitAVG method makes use of the straightforward network split and feature map concatenation strategies to encourage the federated model training an unbiased estimator of the target data distribution. This was achieved by combining the information from multiple feature maps. We evaluate the performance of SplitAVG in comparison to seven cutting-edge federated learning methods by utilizing a centrally hosted training dataset as the baseline for evaluation on a variety of synthetic and real-world federated datasets. We discovered that the performance of models that were trained using any of the comparison federated learning methods significantly decreased as the degrees of data heterogeneity increased. On the other hand, the SplitAVG method achieves results that are comparable to those obtained by the baseline method in all heterogeneous settings. Specifically, it achieves 96.2% of the accuracy and 110.4% of the mean absolute error obtained by the baseline in a diabetic retinopathy binary classification dataset and a bone age prediction dataset, respectively, on highly heterogeneous data partitions. These results are based on the fact that the SplitAVG method is applied to highly heterogeneous data partition We have come to the conclusion that the SplitAVG method is capable of effectively overcoming the performance drops that are caused by the variability in data distributions across institutions. The results of the experiments also indicate that SplitAVG is generalizable to a wide variety of medical imaging tasks and can be modified for use with a variety of base convolutional neural networks (CNNs).