This multi-institutional study investigated the performance of RP using models with pseudo-structures and determined the optimal number of pseudo-structures for RP models. Inter-model comparisons showed no major differences for most dosimetric parameters, as shown in Table 3 and Fig. 4. Compared with the clinical plans, the RP plans showed better dose coverage and OAR sparing. Furthermore, Table 4 indicates that most RP plans were able to achieve the acceptable criteria of the JCOG 1402 protocol. Previous publications evaluating RP models showed that KBP plans exceeded the clinical accepted plan quality at various anatomical sites [7, 16, 30, 31]. Although the models in these studies were registered with multiple original OAR structures, the RP performance of our models with only 2–5 pseudo-structures registered was found to be similar to that reported in previous publications. Thus, our results show that training RP models with pseudo-structures is a simple and effective approach for creating high quality VMAT plans with RP.
The dosimetric parameter results in Table 3 and Fig. 4, and the proportions of plans achieving acceptable criteria in Table 4, indicate that Model_2 using only the PTV and Control_A P showed good dosimetric performance at all institutes. As shown in Table 2, a line objective was commonly used for Control_A P at each institute. The good dosimetric performance of Model_2 is helped by the use of the line objective defined slightly below the estimated dose volume histogram (DVH) lower bound, which helps to drive the optimization towards the best estimated DVH [32]. As the weights for the points on each line objective are all equal, reducing the average dose for large volume structures may be more effective. However, the dose sparing of the femoral head in RP_2 was the worst among the RP models at each institute. This was because the Control_A P pseudo-structure did not impose dose constraints locally on the femoral head. However, the proportions of plans achieving acceptable criteria in the right and left femoral heads were 100% and 98%, respectively. Model_5, in which Control_A P was divided and registered in the model, also showed no dose-reduction advantage in the rectum, bladder, and pelvic bone in comparison with Model_2. Therefore, we conclude that the RP model can perform adequately with the registration of two structures.
The potential benefit of this modeling approach using pseudo-structures is time efficiency. In practice, the correcting of tumor and normal tissue variations through modification of the original plan is hampered by the time-consuming re-planning process, which currently represents the major obstacle for large scale implementation of this strategy [33]. Recently, the use of deformable image registration (DIR) for automatic propagation of structures in ART has been widely investigated. However, registration errors may still exist with DIR, especially for structures that are small and lack contrast with the background, and these registration errors could result in significant dosimetric deviation [34]. Additionally, Nelson et al. reported a total planning time of 207.5 minutes for OAR contouring and optimization, even when implementing DIR in an adaptive plan with the assistance of KBP [15]. As Table 5 shows, the plan created using Model_2 took only 17 minutes for the pseudo-structure creation and optimization process. Acharya et al. reported that the median time for online ART using an MR Linac was 26 minutes, including re-contouring, re-optimization, and patient-specific quality assurance [35]. Therefore, this modeling approach using pseudo-structures should be useful for online ART.
The KBP approach has the advantage that its model can be shared by multiple institutions. Sharing of models is considered to be a good method for reducing variability in planning quality across multiple institutions [17]. In the present study, we were able to create a plan achieving acceptable criteria with a model that was created using only a simple procedure manual. Therefore, the model can be easily shared by creating pseudo-structures at each institute. It was also reported that inter-observer contouring variations have a significant impact on dosimetric and radiobiological outcomes in intensity modulated radiation therapy planning [36]. Reducing the number of structures is useful as a means of homogenizing treatment plan quality across institutions.
The planning quality evaluation in this study was conducted only for cervical cancer patients, and there is a limitation in that the methods in this study cannot cover treatment plans for several PTV dose levels using a simultaneous integrated boost (SIB), such as is performed in head and neck cancer patients; management of the dose gradient around PTVs is more complex with SIB-VMAT plans. In addition, the structures used for the dose evaluation were not considered in this study, but may be defined using automatic segmentation methods.
