ESTRO 38 Abstract book

S221 ESTRO 38

to collapse the Pareto surface plans into a single deliverable plan. The goal is to make the navigated dose an as realistic as possible representation an achievable dose distribution. Material and Methods The multi-criteria optimization (MCO) module of RayStation (RaySearch Laboratories) was augmented with the capability to generate Pareto surfaces composed of VMAT plans represented by control points. All plans were made to satisfy a movement pattern where the MLC leaves sweep back and forth across the target volume as the gantry rotates, defined such that the leaf motion was unidirectional within sub-sectors of an arc. A deliverable plan was, following navigation, constructed by interpolation of the leaf trajectories of the Pareto surface plans. The interpolated trajectories were sampled into control points at equispaced gantry angles and the control points optimized towards minimization of any dose error relative to the navigated dose. Results The new form of MCO was evaluated for three patients (head and neck, lung, and brain with multiple metastases). All patients were planned for treatment with a 360-degree arc. For each patient, a Pareto surface was generated using a fast optimization dose algorithm. A navigated dose satisfying the patient’s clinical goals was then identified and this dose converted to a deliverable plan. Final dose for the deliverable plan was calculated using an accurate dose algorithm and the accuracy between this dose and the navigated dose then assessed. The resulting differences are quantified in Table 1 and summarized qualitatively as DVHs in Figure 1. Table 1. Voxelwise dose differences between the navigated dose and the deliverable plan, with differences corresponding to an improved dose (a dose closer to the prescription level for targets and closer to zero otherwise) truncated to zero.

bladder (low), rectum, PTV, EBV (high).

Results The GAN network was trained over 75 epochs. The average difference between the 3D predicted and clinical dose distributions was 0.1%±2.6% and 0.9%±1.9% for PTV and EBV among 76 patients. For 5 patients, ASEQ was used to generate automated IMRT plans targeting the predicted dose (Figure 2). The average difference between the clinical and ASEQ dose was -1.1%±3.1% and -0.3%±2.4% for PTV and EBV while bladder and rectum were more spared (10.7% and 3.6% on average). Total time per patient for dose prediction and plan optimization was 47 and 109 sec respectively.

Conclusion We demonstrate that conditional GAN networks can be used to accurately predict the 3D dose distribution of prostate patients. Moreover, the GAN output can be directly used into our fast automated optimization as a voxel-by-voxel prescription to generate deliverable plans closely matching the clinical plans. Application of this pipeline can be twofold: the initial dose prediction could assist the radiation therapist during the offline plan generation, while in an MRI-linac setting the ASEQ pipeline can be used for fast plan generation on an online inter/intrafraction basis. We are now evaluating automated plan generation for all test patients and further exploring both the network and ASEQ-specific parameters. PV-0424 Deliverable multi-criteria navigation for VMAT in RayStation R. Bokrantz 1 1 RaySearch Laboratories, Research, Stockholm, Sweden Purpose or Objective To develop and evaluate a method for deliverable multi- criteria VMAT navigation: a form of real-time planning based on exploration of a Pareto surface represented by a set of plans and their linear combinations. Navigation has conventionally relied on Pareto surface representations composed of plans generated by fluence map optimization. Such a representation necessitates a post- processing step where the navigated dose is converted to control points, which can degrade the dose. Deliverable navigation, in contrast, uses plans that are segmented into control points in combination with a control point interpolation technique that after the navigation is used