Fig. 3: T1Gd synthesis and BBB status identification performance on the MR-1 testing set (n = 96 samples).

All analyses were based on 96 independent subjects with no technical replicates. a The receiver operating characteristic (ROC) and precision-recall curves (PRC) for BBB status identification using different methods. b Confusion matrices comparing our CBSI with two direct-identification methods—one using (w) real T1Gd and the other without (w/o). c t-SNE scatter plot of T1Gd features extracted during BBB status identification from real and synthetic images generated by CBSI; orange denotes disrupted and blue intact BBB. d Two representative visual comparisons of T1Gd images generated by the SOTA synthesis models. All evaluations use postprocessed images with pixel values mapped to [0, 255]. Tumor heatmaps within MRI images depict CA-induced enhancement (non-negative component of (T1Gd minus T1)), displayed within 0–150 to highlight subtle differences. Heatmaps below each MRI image show pixel-wise differences between synthetic and real T1Gd (synthetic minus real) within [−150, 150]; warm colors indicate higher positive, while cool colors indicate higher negative differences. e Quantitative evaluation of synthetic T1Gd images across the whole brain, grouped by BBB status (Intact n = 40; Disrupted n = 56). Each data point represents the mean metric from one patient’s brain region (96 independent biological replicates). Violin plots show MAE, SSIM, PSNR, and LPIPS performance. Each violin shows the 25th, 50th (median), and 75th percentiles as dotted lines, with boundaries spanning the full kernel density and all data points overlaid. f As in (e), but with a statistically analyzed tumor region in the form of box plots. Each box shows the median (solid line), the 25th and 75th percentiles (box bounds), and whiskers extending to values within 1.5 times the interquartile range. Statistical significance was assessed using two-sided Wilcoxon signed-rank tests, with exact p-values provided.