Abstract
An intra-planner study has been carried out to compare planning choices among qualified professionals when utilising Pareto-guided automated planning (PGAP) navigation. PGAP was used to calibrate planning goal weights of a protocol-based automatic iterative optimisation automated planning system. Four qualified professionals (Participant A-D) navigated solutions for eight prostate seminal vesicle (PSV) patient cases using PGAP. Plans were based on an existing clinically approved planning protocol containing seven planning goals (PGs). Three PG weights were navigated per plan (rectum $D_{mean}$, bladder $D_{mean}$ and PTV conformality) with all other weights held constant at a value assigned in the original clinically approved protocol. Statistically significant differences were observed between participants for all PG groups except bladder $D_{mean}$. However, dosimetrically the PGAP system mitigated the majority of discrepancies in deviations at the calibration stage with few statistically significant dose-volume metric differences observed, none of which were clinically significant.
Introduction
Radiotherapy planning is a high-stakes balancing act: clinicians must deliver a lethal dose of radiation to a tumor (the Planning Target Volume, or PTV) while sparing adjacent healthy organs (Organs at Risk, or OARs). Traditionally, this was a manual, iterative process prone to "inter-planner variability"—where the quality of a patient's treatment might depend on which specific dosimetrist or physicist was working that day.
Automated planning (AP) aims to solve this, but "black box" automation can lack the nuance of clinical intuition. Pareto-guided automated planning (PGAP) offers a middle ground, allowing experts to navigate a "Pareto front"—a mathematical boundary where one goal cannot be improved without degrading another. This paper investigates the degree of variability in expert choices when using these interactive tools and whether the "human in the loop" reintroduces inconsistency into automated workflows.
Methods
The researchers utilized a framework called EdgeVcc (Experience Driven plan Generation Engine), which uses Protocol-Based Automatic Iterative Optimisation (PBAIO). Four qualified professionals participated: one medical physicist (Participant A), two oncologists (Participants B and D), and one professional planner (Participant C).
The experts planned for eight randomly chosen prostate seminal vesicle (PSV) cases. They were provided a protocol with seven planning goals (PGs) but were only allowed to "navigate" (adjust the weights of) three specific trade-offs: rectum $D_{mean}$, bladder $D_{mean}$, and PTV conformality. All other weights were held constant. The study used Sørensen–Dice coefficients (DiceC) to measure the similarity between the participants' weighting choices and ANOVA/Friedman tests to check for statistical significance in the resulting dose distributions.
Results and Key Insights
Professional Bias and Navigation Styles
The study found that professional background significantly influenced planning priorities. Oncologists (Participants B and D) consistently prioritized sparing the rectum more than the physicist or the professional planner did. This reflects a specific clinical preference to protect the bowel, even if it slightly compromises the conformality of the dose around the tumor. Furthermore, navigation styles varied; Participant D utilized a much wider range of the navigation scale, while Participant B was more conservative.
The System as a Stabilizer
Despite the statistically significant differences in the weights chosen for almost all navigated goals (except bladder $D_{mean}$), the actual dose delivered to the tissues showed very few statistically significant differences. The PGAP system acted as a mathematical constraint that prevented individual preferences from veering into clinically unacceptable territory. Any observed differences in dose were tiny; for instance, the mean difference in the PTV60 $D_{98\%}$ (a measure of dose coverage) between participants was only 0.155 Gy, which is considered clinically negligible in radiotherapy.
Discussion and Limitations
The findings suggest that PGAP successfully buffers human variability while still allowing for clinical input. The high agreement on bladder sparing across all participants suggests that some clinical trade-offs are more settled or less contentious than others (like the rectum vs. PTV conformality trade-off).
The authors acknowledge that the study is limited by its small scale: only four participants and eight patient cases. Additionally, the study focused on a single treatment site (prostate) and a specific software framework (EdgeVcc). Future research could expand this to more complex anatomical sites, such as head and neck cancers, where trade-offs are more numerous and difficult to navigate.
Conclusion
While experts prioritize clinical goals differently based on their professional background, the underlying automated framework in PGAP successfully stabilizes these variations. This results in clinically consistent, high-quality treatment plans, proving that interactive automation can maintain standardization without removing the necessary element of clinical judgment.
Summary updated