Session Item

In automated online adaptive replanning, a common challenge is to create a plan that restores the initially planned dose, mitigating dose degradation effects that arise from changes in patient anatomy, positioning, and tissue densities. This can be formulated as solving a dose restoration optimization problem where one tries to maximize the similarity to the planned dose. However, finding suitable objective functions quantifying this similarity is difficult. For this purpose, we introduce new cross-entropy based objective functions for use in adaptive replanning. By optimizing with these objectives an adapted dose similar to the planned dose is obtained without planner interaction. The viability of the approach is demonstrated on a clinical oropharynx case.

A planning CT and daily fraction kVCTs from a head and neck patient previously treated with a Radixact unit were used to test the proposed approach. Each fraction image was deformably registered to the planning CT and converted into a synthetic CT. OARs were then segmented on the synthetic CT using a commercial deep-learning model, while CTV and PTV margins were copied from the planning to the synthetic CT by rigid registration. For each fraction, cross-entropy based objective functions were applied to the ROIs included in the optimization problem used for the original plan and used in optimization to generate an adapted plan. The cross-entropy objective functions quantify the difference to DVH-metrics of the original plan, as well the voxel-wise dose difference to a reference dose obtained by mapping the original dose to the synthetic CT by deformable registration. This way, the adapted dose mimics the planned dose in both the DVH domain and the spatial domain.

For each fraction, the dose from the adapted plan was compared both against the plan dose on the planning CT, as well as the estimated fraction dose, i.e., the dose obtained if the original plan were to be delivered to the synthetic CT. For most fractions, a decrease in patient volume during the treatment lead to the estimated fraction dose being too high in the high dose target region, and for some fractions the target coverage in the fraction dose was poor. For all fractions, the adapted plan successfully mitigated such dose degradations when present and managed to restore the plan dose in terms of DVH-metrics as well as spatial distribution. The result for one of the fractions can be seen in the attached figures.

The proposed cross-entropy based objective functions are a promising tool for restoring a planned dose on a daily anatomy. They can be integrated in an automated online adaptive replanning workflow and used to mitigate dose degradation effects from variations in patient geometry.