Figure 1. The proposed multi-tissue compartment modeling pipeline for uncovering intratumor vascular heterogeneity.

(a) On the DCE-MRI sequence, tumor region is extracted using a digital mask. Then, pixel time-courses are collected and normalized over time. (b) Pixel time-courses are grouped into clusters with initialization-free multivariate clustering techniques. On the simplex of pixel time-courses, the clusters present at the vertices are identified by a convex analysis of mixtures. (c) Using pure-volume pixels, multi-compartment modeling is performed to estimate tissue-specific flux rate constants and volume transfer constants. (d) Scatter simplex of real DCE-MRI data from an advanced breast cancer. (e) Estimated tissue-specific compartmental time-activity curves: ‘blue’ – plasma input function; ‘red’ – fast flow kinetics; ‘green’ – slow flow kinetics; and example images of the associated local volume transfer constants. (f) Illustrative microscopic images of normal and abnormal vessel architecture (McDonald and Choyke, Nat Med 9, 2003).