ESTRO 2024 - Abstract Book

S4558

Physics - Machine learning models and clinical applications

ESTRO 2024

Among all models, extra trees was the most promising, with a mean absolute error (MAE) of 2.4% on the test set. Figure 1 shows for these models the distribution of the residuals, i.e., the difference between measurements and predictions. Small positive median values were observed, with values of 0.7% and 1.1%, for the models trained with and without sample weights, respectively. At the same time, a tendency to overestimate lower values of the experimental GPR was noticed, even with the use of a weight sampling strategy. A more pronounced difference was displayed by the two classification models. While the model trained without class weights yielded a sensitivity - number of fields with GPR<95% correctly identified - of 0.38 and a specificity - number of fields with GPR>=95% correctly identified - of 0.97, the model trained with class weights showed improved performance, with a sensitivity of 0.60 and a specificity of 0.93. Table 1 summarizes the evaluation metrics for the regression and classification models.

Conclusion: