Figure 4.

Global error analysis from the prediction scenario. The top panel shows model weights determined from the Akaike Information Criterion for the five animals from the prediction scenario. C₁ through C₃ correspond to the different approaches to introduce spatial variation in tumor response, while RTM₁ through RTM₃ correspond to the different RT response models. Models with a local k_p,T and RTM₃ accounted for 54.9% of the ensemble average model. In the bottom panel, the average percent error in tumor volume and the Dice correlation coefficient are reported for the selected model (orange) and the ensemble averaged model (blue) for five animals. (Animals 2 and 4 had insufficient imaging visits for the prediction scenario). The median error for the selected model resulted in less than 16.2% error for all animals. For animal 1 and 6, the ensemble average greatly overestimated the tumor volume with a median error of greater than 100%. A high level of spatial overlap (Dice values greater than 0.60) was observed for the selected model. Statistically significant differences (p < 0.05) between model estimates are indicated by the ‘+’ symbol.