Machine Learning Approach and Model for Predicting Proton Stopping Power Ratio and Other Parameters Using Computed Tomography Images
Purpose: The purpose of this study was to accurately estimate proton stopping power ratio (SPR), relative electron density ρe, effective atomic number (Zeff), and mean excitation energy (I) using one simple robust model and design a machine learning algorithm that will lead to automation. Methods: E...
Saved in:
| Main Author: | |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Wolters Kluwer Medknow Publications
2024-12-01
|
| Series: | Journal of Medical Physics |
| Subjects: | |
| Online Access: | https://journals.lww.com/10.4103/jmp.jmp_120_24 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Purpose:
The purpose of this study was to accurately estimate proton stopping power ratio (SPR), relative electron density ρe, effective atomic number (Zeff), and mean excitation energy (I) using one simple robust model and design a machine learning algorithm that will lead to automation.
Methods:
Empirical relationships between computed tomography (CT) number and SPR, ρe (Zeff) and I were used to formulate a model that predicts all the four parameters using linear attenuation coefficients which can be converted to CT numbers. The results of these models were compared with the results of other existing models. Thirty-three ICRU human tissues were used as modeling data and 12 Gammex inserts as testing data for the machine learning algorithm designed. More ways of tissue classification were introduced to improve accuracy. In the examples, the dual energy methods were implemented using 80 kVp and 150 kVP/Sn.
Results:
The proposed method gave modeling root mean square error (RMSE) near 1% at maximum for the case of SPR and ρe for both single and dual-energy CT approaches considered with modeling RMSE of 0.32% for ρe and 0.38% for SPR as modeling RMSE with room for improvement (this can be done by adjusting the model number of terms as well as the parameters). The method was able to achieve modeling RMSE of 1.11% for I and 1.66% for Zeff. The mean error for all the estimated quantities was near 0.00%. In most cases, the proposed method has lower testing RMSE and mean error compare to the other methods presented in the study.
Conclusion:
The proposed method proves to be more flexible and robust among all presented methods since it has lower testing error in most cases and can be improved based on data using the machine learning algorithm. The algorithm can also improve estimation by adjusting the model as well as aid in automation and it’s easy to implement. |
|---|---|
| ISSN: | 0971-6203 1998-3913 |