Effective Dose Estimation in Computed Tomography by Machine Learning

Area of Science:

• Medical Imaging
• Radiological Physics
• Machine Learning in Healthcare

Background:

• Computed tomography (CT) scans are essential diagnostic tools, but patient radiation exposure requires careful management.
• Accurate estimation of effective dose (E) is crucial for ensuring patient safety in CT imaging.
• Current methods for dose estimation can be complex or require specialized software.

Purpose of the Study:

• To develop and evaluate machine learning models for predicting effective dose (E) from CT scans.
• To estimate E using only patient and CT acquisition parameters, bypassing the need for dose-tracking software.
• To compare the performance of machine learning algorithms against traditional dose estimation methods.

Main Methods:

• Utilized a dataset of 69,037 CT acquisitions with dose-tracking software (DTS) for training and validation.
• Trained and optimized various machine learning algorithms, including random forest, neural networks, and support vector machines.
• Evaluated model performance using Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), and R-squared, including testing on an external dataset.

Main Results:

• The random forest regressor achieved the best performance, with MAE of 0.416 mSv and MAPE of 7% on the test set.
• Machine learning models significantly outperformed traditional methods like k-factors (MAE: 2.06 mSv) and multiple linear regression (MAE: 0.98 mSv).
• The random forest model demonstrated strong generalization on an external dataset, yielding an MAE of 0.215 mSv and MAPE of 7.1%.

Conclusions:

• Machine learning models can accurately estimate effective dose (E) in CT imaging using readily available patient and scanner parameters.
• This approach provides a viable alternative to complex dose-tracking software for rapid and reliable dose assessment.
• The findings support the integration of machine learning for improved radiological safety and dose management in clinical practice.