Prediction of radiation pneumonitis with machine learning using 4D-CT based dose-function features.

Area of Science:

• Medical Physics
• Radiation Oncology
• Radiotherapy Imaging

Background:

• Radiation pneumonitis (RP) is a significant side effect of thoracic radiotherapy.
• Accurate prediction of RP is crucial for optimizing treatment plans and patient outcomes.
• Current prediction models often lack sufficient predictive power.

Purpose of the Study:

• To evaluate the efficacy of combining dose-volume histogram (DVH) and dose-function histogram (DFH) features for predicting radiation pneumonitis (RP).
• To develop and validate a multivariate prediction model using machine learning techniques.
• To assess the added value of functional image-derived DFH features over traditional DVH features.

Main Methods:

• Development of multivariate prediction models using kernel-based support vector machine (SVM) machine learning.
• Calculation of DVH and DFH features from functional images derived from 4-dimensional computed tomography (4D-CT) and deformable image registration (DIR).
• Inclusion of Hounsfield unit (HU) and Jacobian metrics for DFH feature extraction in 85 non-small cell lung cancer patients.

Main Results:

• A prediction model incorporating patient clinical features and DVH features achieved a median area under the curve (AUC) of 0.58.
• Adding HU metric DFH features improved the AUC to 0.73.
• Incorporating Jacobian metric DFH features resulted in an AUC of 0.68.
• 21 out of 85 patients (24.7%) developed RP grade ≥ 2.

Conclusions:

• Predictive models for RP incorporating DFH features were successfully developed using kernel-based SVM.
• The simultaneous use of DVH and DFH features derived from 4D-CT and DIR significantly enhances the predictive accuracy of RP.
• Functional image-guided radiotherapy benefits from incorporating both DVH and DFH features for improved treatment planning.