Fig. 2: Efficacy of HSL vs. classical methods.
From: Addressing data heterogeneity in distributed medical imaging with heterosync learning
A1 Data distribution across the 7 nodes for feature skew. A2 AUC box plots of model testing efficacy on each node (n = 440 per testing set) for various learning methods for feature skew. B1 Data distribution in label distribution gradients of the 2 nodes for label skew. B2 AUC line charts of model testing efficacy of the 2 nodes (n = 1191 per testing set) for label skew. C1 Data distribution in quantity distribution gradients of the 2 nodes for quantity skew. C2 AUC line charts of model testing efficacy of the Node 1 (testing set sizes: n = 1301, 867, 520, 372, 289, 236, 123, 63, 42, and 31 per gradient) and Node 2 (testing set sizes: n = 1301, 1735, 2083, 2232, 2314, 2367, 2479, 2540, 2560, and 2571 per gradient) for quantity skew. D1 Data distribution in the 5 nodes for combined heterogeneity. D2 AUC box plots of model testing efficacy across nodes under combined heterogeneity (testing set sizes: n = 938, 665, 60, 60, and 4279). Box plots in A2, D2 show the median (central line), the 25th and 75th percentiles (box bounds), and the minimum and maximum values excluding outliers (whiskers). Outliers are shown as individual points. Each bar represents the results of five repeated experiments for a learning method. Source data are provided as a Source data file. Figures aregenerated using Python and subsequently composited in Adobe Illustrator 2023.
