Automated Analysis of CBCT Frequency Using a Polygon-Based Training Model and Random Forest Classification
Abstract
Purpose
Cone-beam CT (CBCT) is routinely used for image-guided radiation therapy (IGRT); however, in the optimal frequency of CBCT varies across patients and treatment sites, affecting imaging dose, workflow efficiency, and clinical consistency. This study aims to develop an automated machine-learning–based framework to evaluate CBCT acquisition frequency and to assess whether CBCT from the first-three treatment fractions can achieve comparable clinical objectives to daily CBCT.
Methods
CBCT data from 79 lung cancer patients treated under an institutional IGRT protocol were analyzed. The protocol employs daily CBCT for the first-three fractions followed by weekly CBCT for soft-tissue alignment, with daily orthogonal kV imaging for bony anatomy. Patient eligibility for reduced CBCT frequency (eGroup) was determined based on CBCT positional match criteria. A polygon-based training approach was used to encode spatial features associated with CBCT acquisition patterns. The CBCT frequency were classified using a Random Forest classification algorithm, an ensemble tree-based learning method that aggregates predictions from multiple decision trees through majority voting, providing robustness against overfitting and noise. Model performance was evaluated using standard classification metrics, including specificity and sensitivity curve.
Results
The sensitivity and specificity curves with λ from 40% to 70% with step size of 5%, as well as Δ values of 3, 4, and 5 mm are shown in Figure 4. When λ equals 50%, sensitivities for identifying a protocol-ineligible group (iGroup) patient were 0.83, 0.83, and 0.83, specificities for identifying an eGroup patient were 1.0, 1.0, and 1.0, respectively.
Conclusion
The proposed polygon-based feature representation combined with Random Forest classification provides an accurate and interpretable framework for CBCT frequency analysis. The results demonstrate that CBCT data from the first three treatment fractions can achieve clinical objectives comparable to daily CBCT, supporting data-driven optimization of IGRT imaging protocols.