Federated orthogonal learning for detection of liver lesions from multi-phase contrast-enhanced CT images

Multi-phase contrast-enhanced CT (CECT) scans are often scattered across multiple institutions and contain incomplete phase in parts of institutions due to the strict data-protection regulations and the disparity of phase integrality. While federated learning (FL) enables training a privacy-preserving model collaboratively across institutions, it often suffers from significant performance degradation for liver lesion segmentation caused by heterogeneity in different institutions. To tackle the challenges, we present FedOG to guide a deep convolutional neural network collaboratively segment liver lesions for minimizing interference on 3,668 CECT multi-phase CECT scans from five different institutions. Specifically, FedOG adjusted the gradients from local models trained with incomplete phases of CECTs via orthogonal gradient decomposition to alleviate the interference. During each adjustment, the optimal gradient for updating the global model is determined by Bayesian optimization. Experiment results have shown that FedOG improves the Dice score by 1.67% on a large real-world clinical dataset, 1.13%, and 3.03% on two publicly available datasets. We anticipate our study will enable a heterogeneity-robust, search-efficient, and privacy-preserving federated training framework using multi-phase CECT. We also found out that FedOG is especially beneficial for underdeveloped regions where institutions often have missing or low-quality phases of multi-phase CECT scans.