Author: Andres Portocarrero Bonifaz, Ian Schreiber 👨🔬
Affiliation: CARTI Cancer Center 🌍
Purpose: To explore how calculation grid resolution, along with other planning factors, affects head and neck dose calculation accuracy and contributes to potential discrepancies in the Eclipse Treatment Planning System, and to construct a machine learning model to predict when such discrepancies are likely.
Methods: A total of 185 patients with head and neck lesions were included, resulting in 289 plans. Each plan was recalculated at grid resolutions ranging from 1 mm to 3 mm in increments of 0.5 mm, using both the AAA and Acuros algorithms. Dose-volume histogram statistics were collected for all relevant structures, producing 30,396 data points. Factors analyzed included organ-at-risk (OAR) and target hot-spot, centroid, and minimum OAR–planning target volume (PTV) separation, sampling coverage, and OAR and PTV volumes. An F–beta–scored random forest model was trained to predict whether a structure’s dose would deviate from the reference calculation at 1 mm grid resolution. Model hyperparameters were optimized via Bayesian search, and five-fold cross validation was performed to reduce variance. Additionally, variable importances were calculated. Finally, a single cross-validated decision tree was also constructed to offer an alternative, interpretable view of the main factors contributing to these discrepancies.
Results: The random forest achieved a mean F1 score of 0.718 (95% CI: 0.683–0.751), recall of 0.817 (95% CI: 0.776–0.852), and precision of 0.640 (95% CI: 0.599–0.684). The five most relevant parameters in the random forest model were OAR hotspot, target-to-OAR centroid separation, minimum target-to-OAR separation, and sampling coverages for target and OAR. The decision tree supported the hypothesis that large dose grids and small OAR volumes are correlated with dose discrepancies.
Conclusion: The probability of clinically unacceptable dosimetric discrepancies depends on key planning factors that were identified in this study. The interpretable machine learning model adequately predicted situations of dosimetric discrepancy.