ESTRO 2024 - Abstract Book

S4450

Physics - Machine learning models and clinical applications

ESTRO 2024

The interrelation between plan and complexity metrics was examined, as well as the analysis of effective differentiation of QD by these metrics. Three models for QD forecasting based on discriminant analysis (DA), random decision forest (RDF) methods, and the hybrid model were built and evaluated.

Results:

Most of the interrelations between plan and complexity metrics were small, fair or moderate. The exception is correlations of the join function (J) with the average number of monitor units per control point (aMU/CP) (R=0.893) and the beam aperture (BA) with planning target volume (PTV) (R=897). While many metrics (e.g., BA, aMU/CP, or J) allow for the effective separation of the QD from each other, the study shows that predicting values of the QD is possible only through multi-component forecasting models. Analysis of the efficacy of the DA, RDF and hybrid models shows that the hybrid model is the most accurate (0.894 related to 0.875 for the RDF model and 0.550 for the DA model).

Conclusion:

Forecasting the QD results is possible only based on the multi-component models. Of the three models explored in this study, the hybrid model, which uses DA methods to predict red QD and RDF methods to predict green and yellow QDs, is the most promising (Table 1).

Table 1. Descriptive statistics for the models of discriminant analysis (DA), random decision forest (RDF) and hybrid model, and the values of sensitivity and specificity from the model related to specified qualitative descriptors (green, yellow, red).

Keywords: complexity, plan metrics, PSQA