Automated prediction of dosimetric eligibility for hypofractionated prostate radiotherapy

Abstract Clinical implementation of hypofractionated prostate radiotherapy (PROFIT trial, NCT003046759) represents an opportunity to significantly reduce the burden of treatment on the patient and clinic. However, efficacy was only demonstrated among the patient demographic who could meet the trial dose constraints and so it is necessary to emulate this triage step in clinical practice. The purpose of this study was to build a convenient tool to address the challenge of determining patient eligibility for hypofractionated treatment within the clinic. The tool was implemented within the EclipseTM treatment planning system using the scripting environment. Prior to planning a new case, the script computes and displays in a plot the fractional overlap of rectal and bladder wall with the planning target volume. Radial decision boundaries separate the plot into three zones and the new case is then classified as “feasible”, “uncertain”, or “not feasible”. The radial decision boundaries were derived from a retrospective analysis of the overlap values and dosimetric eligibility of 150 patients with intermediate risk prostate cancer. Two‐fold cross validation with repetitions demonstrated an average prediction accuracy of over 90%. The tool has been integrated into our clinical planning workflow to enable early identification of the need for planning consults and rapid a‐priori determination of dosimetric eligibility for hypofractionated radiotherapy. The tool can be readily adopted by other centres since the underlying metrics can be evaluated without scripting if desired.


| INTRODUCTION
Dose-escalated external beam radiotherapy (e.g., 7400-7800 cGy) using intensity modulated radiotherapy (RT) has been shown to improve biochemical control among patients with intermediate-risk prostate cancer 1 and so has become the standard of care. However, this approach involves prolonged treatment lasting 7-8 weeks and so increases the burden of treatment on both the patient and the clinic. patients. At a median follow-up time of 5 yr, the hypofractionated treatment course was found to be non-inferior to conventional treatment with respect to disease control and normal tissue toxicity among the 1206 patients enrolled in the study. 2 Clinical implementation of PROFIT-trial hypofractionation represents an opportunity to significantly reduce the burden of treatment on the patient and clinic. However, the PROFIT trial employed strict dose constraints for the rectum and bladder in order to safeguard against the potentially increased risk of normal tissue toxicity in the hypofractionated study arm. Patients were excluded from the trial if they could not meet these constraints. Consequently, non-inferiority was only demonstrated among the patient demographic whose anatomy was well-suited towards meeting constraints and so it is necessary to emulate this triage step in clinical practice.
The purpose of this study was to build an automated decisionsupport tool to address the challenge of determining patient eligibility for hypofractionated treatment within the clinic. We propose an anatomy-based metric to provide a-priori prediction of whether constraints can be met for a given patient. The predictive efficacy of the metric is evaluated using 2-fold cross validation. We integrate the tool into the Eclipse treatment planning system (TPS) via scripting to provide convenient decision-support to the clinic.  (Table 1). Accordingly, patients were considered to be "dosimetrically eligible" if their plans could meet the accepted constraints. If these constraints could not be met for a given patient by two physicists, the patient was considered to be "dosimetrically ineligible" for hypofractionation.

2.B | Overlap metric
Initial local experience with generating PROFIT trial hypofractionated treatment plans found that rectal and bladder wall D30 constraints were often unmet for challenging cases. In the presence of high PTV-OAR overlap, target coverage or OAR dose must necessarily be compromised since OARs are to receive much less than the prescription dose. In particular, it might be expected that constraints cannot be met if the PTV overlaps more than 30% of the rectal or bladder wall since 30% of these OARs are required to receive less than 4710 cGy. The fractional overlap of PTV with the rectal and bladder wall may therefore serve as a strong anatomical indicator for the ability to meet PROFIT-trial dose constraints.
Let f RW and f BW represent the fraction of the rectal and bladder wall that overlap with the PTV respectively. We propose that f RW and f BW can be added in quadrature to obtain a simple single-valued overlap metric which can be used to differentiate between eligible and ineligible patients. This metric was evaluated for each of the 150 patients and then used for prediction of dosimetric eligibility and assessment of planning difficulty.

2.C | Evaluation of predictive efficacy
Prediction of dosimetric eligbility was performed using a simple threshold technique where a patient would be classified as dosimetrically ineligible if their overlap metric was greater than or equal to a single threshold value. This threshold must be determined by searching for the cut-off value which maximizes prediction of eligibility among a training dataset.
In this study, two-fold cross-validation is employed to obtain this threshold and evaluate its predictive efficacy. Training and testing sets included equal proportions of dosimetrically eligible and ineligible patients. The two-fold cross-validation was repeated 1000 times with different randomly selected training and testing sets for each repetition. The average classification threshold, sensitivity, specificity and accuracy were then computed. We also performed three additional cross validations (with repetitions), one of which used only f BW for prediction, another used only f RW for prediction and the final cross validation investigated the use of the PTV volume for prediction.

2.D | Prediction tool
The prediction tool was created using the Eclipse TM scripting envi-   If the point falls within the green zone as in Fig. 2(b), the new case is considered to be "feasible" since constraints could be met for 100% of patients in this study whose data points were located there  Table 2).
This tool has now been integrated into our local clinical planning workflow. The script is run after contours are generated but prior to treatment planning. Significant effort is expended to meet constraints for any difficult cases located within the green zone (e.g., when a case is close to the green-yellow boundary) since the prior knowledge from this study suggests that this is feasible. If a new case is located in the yellow zone, planning is still attempted but inter-disciplinary discussions regarding planning feasibility are initiated in parallel since it may not be possible to meet constraints for these cases. Finally, hypofractionated treatment planning is not attempted for cases located in the red zone. The planner immediately notifies the radiation oncologist that hypofractionation is "not feasible" and the patient prescription is reverted to a conventional 78 Gy in 39 fraction regimen.  observed ineligibility rate at the planning stage (e.g., exclusion based on seminal vesicle involvement). The overall benefit of the tool reduces as the ineligibility rate decreases since there are fewer infeasible cases for which planning can be avoided. However, the combined overlap metric is still expected to be predictive within this context due to the direct relationship between OAR-PTV fractional overlap and D30 constraints.
In the future, treatment planning may be fully automated for many sites. 3 This would supersede the utility of the tool presented here since there would be no cost associated with generating treatment plans. However, in the medium term, many clinics may not have access to robust automated treatment planning and can therefore benefit from simple and effective treatment planning tools that require no specialized software or expertise.