Evaluation of an automated knowledge based treatment planning system for head and neck

Intensity modulated radiotherapy (IMRT) and volumetric modulated radiotherapy (VMAT) have been used for more than a decade and are now standard techniques for external beam radiotherapy treatment (RT). However, the inverse and computer-based planning approach involves multiple manual steps, which might influence the plan quality and consistency: planning objectives and constraints need to be manually adapted to the patient individual anatomy and tumor location, size and shape [1]. Additional help structures are frequently defined to further individualize and optimize the treatment plan on a patient individual basis resulting in an iterative process of IMRT and VMAT plan generation. Furthermore, the TPS operator needs to have profound knowledge and experience about the limitations of the treatment planning system and techniques, translating this into a prediction of the dose distribution, which can be achieved in each individual case. This method of manual optimization is time consuming, especially for complex cases such as head and neck carcinoma where multiple target dose levels are defined in close proximity to OAR’s.

In order to improve the overall plan quality and consistency, and to decrease the time required for planning, semi-automated planning algorithms have been developed [2, 3]. In the present study, we have evaluated a fully automated treatment planning system (TPS), Pinnacle Auto-Planning (Philips Radiation Oncology Systems), which uses an iterative algorithm based approach to automatically adapt objectives, constraints and dose shaping contours during the optimization process to achieve clinical goals. It was the aim of this study to evaluate Auto-Planning in a large cohort of head and neck cancer patients: treatment plans were compared with historical, clinically accepted VMAT treatment plans generated by Eclipse RapidArc planning (Varian Medical System, Palo Alto, CA). Additionally, time required for manual vs. Auto-Planning was compared.

OAR objectives and priorities used for Auto-Planning are listed in Table 1. These settings were used for the optimization of all head-and-neck cases planned with Auto-Planning. No individualization was performed.

In this comparative study, plans manually optimized with Eclipse were compared to plans automatically optimized with Pinnacle. Both systems are using different optimizer algorithms which potentially influence change the results. Nevertheless, planning challenges involving experienced operators have shown that both systems were able to achieve similar dose distribution [5, 6]. Therefore, we do not believe that the results are biased due to the use of two different optimization algorithms.

Results of our study show that Pinnacle Auto-Planning, an automated iterative planning algorithm, achieved highly consistent treatment plans: target coverage and homogeneity were significantly improved at the expense of dose conformity compared to clinically accepted historical plans. However, dose differences for all target volumes were small and are unlikely to be clinically relevant. Additionally, Auto-planning achieved a significant reduction of all OAR dose. Most importantly, Auto-Planning was able to reduce the dose to at least one OAR by more than 20 % in 64 % of the cases.

This finding of improved OAR sparing by a potentially clinically relevant amount could be explained by sub-optimal constraints being used in the historical manual planning cohort. Plan optimization for head and neck is one of the most complex scenarios due to the complexity of the tumor shape, dose levels and location. Targets with multiple dose levels are defined in close proximity to or as overlapping OARs making the optimization a complex and time consuming process. The result will strongly depend on the skill set of the TPS Operator when setting the correct help structures, optimization objectives and constraints.

In 19 % of the plans, the mean dose to one of the parotids was reduced by more than 20 % with Auto-Planning. This reduction of the mean dose is expected to be clinically relevant and reduce xerostomia [7]. The normal tissue complication probability for the parotid glands reported by Eisbruch et al. is described as a continuous and monotonously increasing function of the mean dose with a steep gradient between 20Gy and 30Gy [8]. Any significant reduction of the mean dose in this range may thus be of clinical benefit. The mean parotid dose increased with Auto-Planning for clinical plans having a dose larger than 35Gy. For these cases, the targets were largely overlapping with the parotid. If this overlap is larger than 25 and 50 %, the priority was automatically reduced to medium or low with Auto-Planning. This could explain the higher mean parotid dose with Auto-Planning in cases where a mean dose greater than 35Gy was observed.

A significant correlation between the mean swallowing muscle dose and complications such as dysphagia has been observed [9]. A steep dose-effect relationship, with an increase of the probability of dysphagia of 19 % with every additional 10Gy, was established [10]. We observed that Auto-Planning could be clinically beneficial for patients in respect to the dose delivered to the swallowing muscles. Indeed, in 20 % of the cases, Auto-Planning was able to reduce the swallowing muscles mean dose by more than 20 % which correspond to a dose reduction of 10.7Gy.

Highly consistent and improved plan quality in a relevant proportion of the cases compared to our historical plans was not only observed in DVH parameters. Indeed, radiation oncologists specialized in head and neck cancer preferred Auto-Planning plans compared to the clinical plans in 80 % of the cases; however they simultaneously stated that results were close and both planning methods achieved clinically acceptable plans in all patients. This result is even more surprising as only one single Auto-Planning head and neck model—pre-set of optimization parameters—was used in Auto-planning and only one single optimization cycle was allowed. One could therefore speculate that further optimization of the Auto-Planning approach might be possible by 1) more customized optimization models and by 2) the use of more than one optimization cycle. By increasing the number of optimization cycles, the planner would be able to change the targets and/or OARs objectives in order to modify the dose distribution previously calculated. This would lead to a more patient specific dose distribution. Further analyses are ongoing to test these hypotheses. Besides potentially improved outcome on a patient-individual level and on an institutional level, such automated planning approaches are considered to be especially useful within the context of multi-center trials to standardize and homogenize planning and plan quality.

Other automated planning approaches have been discussed elsewhere [9, 10]. One method is based on a previously generated site-specific plan library: the individual patient is compared to the patient population in library of plans and similarity measures are used for selection of planning optimization parameters [2, 9, 11]. This compares patient-individual DVH parameters to the database and estimates plan quality in comparison to an overall patient population. However, the application of the library is restricted to patients with identical or very similar planning objectives (number of fractions, target and OAR doses) [9]. Also, a larger number of plans are required for de-novo generation a robust library, which is challenging for smaller institutions and/or rare indications. Finally, since the knowledge of the previous plans is used to generate the objectives, the newly generated plan quality inevitably depends on the quality of the plans building the database [11]. Non-optimal plans entered in the database may degrade results with the plan library approach.

Multicriteria optimization (MCO) approaches [6‐8] have been proposed for real-time assessment of the trade-off between different clinical goals. MCO allows to navigate towards a Pareto optimal plan in a database of automatically generated plans and thereby balance between targets and between organs at risk. One limitation of the MCO approach is the increased time to create the library of treatment plans before dose evaluation can be performed, although this can be overcome by faster dose calculations [12]. The MCO approach also requires significant individual user experience and interaction to determine the Pareto optimal plan, which may become especially challenging in head and neck cases were multiple targets and organs at risk then need to be balanced. In addition, if the MCO generated library of plans is fluence based, the Pareto optimal plan will not be deliverable, and further fluence segmentation is required.

In conclusion, our study shows that the Pinnacle Auto-Planning approach achieves highly consistent treatment plans for complex head-and-neck cancer patients: dose distributions of the target volumes were similar between the automated planning approach and manual planning. OAR sparing was significantly improved using Auto-Planning and potentially relevant improved OAR sparing was observed in more than half of the patients. Furthermore, experienced radiation oncologists preferred the dose distributions from Auto-Planning in comparison to those from the manually generated plans in 80 % of the cases. Human resources required for treatment planning was substantially reduced and made independently from the experience of the treatment planner.