Session Item

The inverse optimization problem of intensity modulated RT has a highly degenerate solution space. Several RT plan designs can produce similar dose distributions which may differ greatly in complexity. Plans with extreme complexity are associated to higher uncertainties and worse patient-specific QA (PSQA) results, whereas under-optimized plans might be less complex but of poor quality. Machine Learning (ML) with a Lean Six Sigma Methodology (LSSM) was implemented to avoid outlier complexities in RT plans and reduce the risk of PSQA failures.

The five DMAIC (Define, Measure, Analyse, Improve, Control) LSSM steps were applied. Define: The RT plan optimization can produce an unnecessary modulation of a linac’s machine parameters, generating a suboptimal delivery to the patient. Measure: Ten complexity metrics were computed for all VMAT plans delivered at our Institution during 2013-2021. Analyse: The distributions of the complexity metrics were examined and stratified by treatment site. Improbable values of complexity were defined as below the 5th- or above the 95th-percentile of the historical distributions, corresponding to either under-optimized or extremely complex plans, respectively. Improve: XGBoost, a tree-based ensemble ML model was trained to predict gamma passing rates (GPR) at 3%/1 mm and absolute dose from the complexity of each arc. A decision support system (DSS) tool was developed for the Eclipse TPS. After each optimization, the complexity metrics and the ML model GPR predictions are shown directly in the TPS. As a visual aid, the complexity metrics with out of range values are flagged and, in case, the plan is re-optimized. Control: The DSS tool was introduced in the clinic on 22nd August 2022, and follow up results were checked after two months.

28424 retrospective VMAT plans (70197 arc) and 211 prospective VMAT plans (525 arcs) were analysed. With the DSS, 48 arcs (9.1%) were found to have more than 5 complexity metrics out of range and 50 arcs (9.5%) were flagged as potential PSQA failures by the ML model, i.e., GPR <92.5% (Figure 1). Corrective actions were taken by either re-optimizing the RT plan or changing the treatment machine. Overall, extreme cases reduced over the Control period. Table 1 reports the percentiles of the distributions of Q1Gap, MeanTGI and 1-MCS metrics, which characterize each VMAT arc in terms of beam aperture, tongue-and-groove effect, and MLC modulation, respectively. The values are shown for representative treatment sites, i.e., H&N, Thorax SBRT, Abdomen SBRT, and Gastro-urinary (GU). For the H&N, Abdomen SBRT, and GU, the 5th-percentile of the MeanTGI increased from 0.29, 0.19, 0.27 to 0.34, 0.26, 0.33, while the 95th-percentile reduced from 0.58, 0.52, 0.57 to 0.53, 0.51, 0.50, respectively.

ML with LSSM was implemented in RT clinical practice, allowing to reduce outlier complexities of RT plans, reducing the risk of PSQA failure and streamlining the RT workflow.