Modeling late rectal toxicities based on a parameterized representation of the 3D dose distribution

Area of Science:

• Radiation Oncology
• Medical Physics
• Biostatistics

Background:

• Traditional Normal Tissue Complication Probability (NTCP) models often use dose-volume histograms (DVHs), which can lose spatial information and have correlated bins.
• Existing complex models may lack physical interpretability and the ability to predict probabilities.

Purpose of the Study:

• To develop and validate a novel parameterized representation of 3D rectal dose distribution for predicting late rectal toxicity.
• To compare the predictive performance of this new model against traditional DVH-based NTCP models.

Main Methods:

• A parameterized representation of the 3D dose distribution to the rectal wall was created, incorporating geometrical features like eccentricity, lateral, and longitudinal extent.
• A nonlinear kernel-based probabilistic model was employed to predict late rectal toxicity using the parameterized dose data.
• Model performance was assessed using data from the MRC RT01 trial, evaluating endpoints such as rectal bleeding, loose stools, and a global toxicity score.

Main Results:

• NTCP models using parameterized geometrical and volumetric measures achieved Areas Under the Curve (AUCs) of 0.66 (rectal bleeding), 0.63 (loose stools), and 0.67 (global toxicity).
• In contrast, DVH-based NTCP models yielded lower AUCs of 0.59 for all three endpoints.
• The study identified simple rules linking specific 3D dose patterns to a lower risk of rectal complications.

Conclusions:

• Parameterized 3D dose representations provide interpretable, low-dimensional, and nonlinear NTCP models with superior predictive power compared to standard DVHs.
• This novel approach enhances the ability to predict rectal toxicity and identify favorable 3D dose patterns in radiation therapy planning.