Fig. 3: Analysis of the fluorophores distribution in vessels after IV injection of TRITC-dextran and transcardial perfusion with AF647-lectin.

a–c 3D rendering of high-magnification confocal Z-stacks of GBM vasculature depicting different regions of vascular tree with respect to fluorophores distribution. Green arrows mark non-perfused vessels aligned solely with signal from the remaining TRITC-dextran. Red arrows mark fully perfused vessels as determined by the lack of signal from TRITC-dextran and the presence of signal from AF647-lectin. Blue arrows mark annotation of underperfused transition zones. d Vasculature represented by immunostaining of collagen IV in the same region. e Overlay of images in a–d. Coloring of the vessels segments follows the same logic as for arrows (a–c). Scale bars, 10 µm. f Diagram illustrating the principles of manual annotation of differently labeled vascular segments for training of the machine learning algorithm. Quantification of the mean pixel intensity of the dextran signal in non-perfused and underperfused vessels in GBM (g) (P < 0.0001, n = 30 mice, Wilcoxon signed-rank test) and in contralateral hemisphere (h) (P < 0.0001, n = 30 mice, Wilcoxon signed-rank test). Data are presented as mean ± SD. Representative images of vessel segmentation performed by the trained model overlaid with original image (i–k) and separately (l). Segmentation color-coding: magenta—non-perfused vessels, cyan—underperfused vessels, and orange—perfused vessels. Scale bars, 30 µm. m–p Quantification of the degree of transcardial perfusion in GBM vessels and vessels from tumor-free contralateral hemisphere. This was studied in animals bearing tumors arising from the G01 (m, n = 10), G06 (n, n =10), and U87 (o, n = 10). p Summary of data from all tumor models (n = 30). Proportions of vasculature classes in tumors are significantly different from those in contralateral hemisphere for all groups (P < 0.0001, logistic regression after arcsin conversion of percentages), differences in proportions of vasculature classes in the tumors between different GBM models are nonsignificant (P > 0.3, logistic regression after arcsin conversion of percentages). More detailed representation of the data in m–p with individual values displayed can be found in Supplementary Fig. 1.