Machine-Learning Prediction of the Gel Dosimetry Calibration Curve Equation Based On the Non-Irradiated Gel Magnetic Resonance Imaging
K Tozzi, J Lizar, J Pavoni*, Department of Physics, Faculty of Philosophy, Sciences and Letters at Ribeirao Preto, University of Sao Paulo, Ribeirao Preto, Sao Paulo - BR
Presentations
PO-GePV-T-102 (Sunday, 7/10/2022) [Eastern Time (GMT-4)]
ePoster Forums
Purpose: To develop regression algorithms for predicting the linear equation of the gel dosimetry calibration curve based on machine learning.
Methods: 50 multi-spin echo magnetic resonance imagings containing MAGIC-f gel vials were used. In each image, a non-irradiated gel vial and other vials irradiated with different doses between 0 and 10 Gy are shown. The calibration curves (dose versus R2) were determined for each image, and by applying a linear fitting to the data, their slope and intercept were achieved. The radiomics features of the non-irradiated vials were extracted from these images using Pyradiomics. A set of decision trees (forest) was used to build the regression model that was fed with radiomics data and the data of the corresponding calibration curves. Two models were developed, one to find the slope value and the other for the intercept. The selection of the most important features for the target and the forest hyperparameter tuning was also made. The data was divided into 70% for training and 30% for testing the model. The model´s performance was evaluated based on the mean square error (MSE) and the mean absolute error (MAE).
Results: 9 features were selected as most important for both models. For the final models, the MSE for the slope and intercept were 4.5 x 10-4 and 6.7 x 10-4, while the MAE for them was 0.019 and 0.02, respectively. Within this analysis, 80% of the data was incorporated into an uncertainty deviation of ± 0.03 while 100% was into ± 0.04 for Slope and 87% of the data was included into ± 0.04 for Intercept in both metrics.
Conclusion: The algorithms developed proved to be satisfactory for the prediction of both coefficients. Thus, is possible to reduce the time involved in the process of gel dosimetry calibration by using this methodology.
Funding Support, Disclosures, and Conflict of Interest: This work was supported by grant 2021/02254-6, Sao Paulo Research Foundation (FAPESP)
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email