Integration of the T-GEP–based MetSig with FDG-PET radiomics defines a RadSig significantly associated with outcome in DLBCL. (A) Graph illustrating radiomic features considered not robust and excluded because of significantly different values according to scanner model, frame duration, time between injection and acquisition, and glucose level. The x-axis shows the parameters and the corresponding P values in negative log₁₀ scale are illustrated on the y-axis. Each circle on the plot represents a single radiomic feature. (B) Correlation matrix heatmap of radiomic features displaying the Spearman correlation coefficient between each pair of radiomic features; radiomic features were reordered by unsupervised hierarchical clustering for visualizing highly intracorrelated features. Five clusters of radiomic features were generated (the red blocks along the diagonal indicate high intracluster correlation, blue squares indicate negative correlation, and red squares indicate positive correlation). (C) A circos plot showing correlation between the MetSig and radiomic features. Only radiomic features with a significant correlation (P < .05) with the MetSig are shown in the circos plot. (D) Heatmap representing the 4 informative radiomic features composing the RadSig shown as rows, and patients with DLBCL samples shown as columns in the discovery cohort. (E) Box plot graph depicting MetSig ratio values in the RadSig-low and -high patient subgroups. P value was calculated with the Mann-Whitney Wilcoxon test. (F) PFS of the discovery cohort according to the RadSig status in RadSig-low vs RadSig-high patient subsets. P value was calculated with the log-rank test. GLCM, gray level coorcuurence matrix; NGLDM, neighboring gray level dependence matrix.