2222 - Permuted Learning Framework for Resimulation Decision Support in Head and Neck IMPT Surface Mapping with Immobilization Device Masking

Purpose/Objective(s):

Daily image guidance for head and neck intensity-modulated proton therapy (IMPT) is challenging due to large target volumes and anatomical changes affecting dose distribution. This study aimed to develop a novel algorithm for immobilization device removal and objective surface mapping, integrated with a permuted learning framework, to predict resimulation(resim) events and enhance decision support.

Materials/Methods:

We collected CT data for across 8475 beams over 3183 fractions from 160 head and neck IMPT patients. An automated algorithm was developed to remove immobilization devices from head and neck CTs using binary image processing, isolating the largest inner contour containing the centroid and adjacent to the patient exterior. Processed contours were reconstructed into 3D surfaces using Poisson reconstruction to calculate surface mapping (SM) scores by comparing daily CBCTs to planning CTs.

A permuted learning framework systematically shifted ground truth resim decisions by ±1 fraction to account for clinical variability. Temporal patterns of SM score deviations along beam projections were used as input features. XGBoost classifiers with class imbalance correction evaluated each permutation. The optimal model was selected by maximizing area under the curve (AUC) while maintaining positive recall above 90%, ensuring high sensitivity to resim events. Statistical analyses, including interaction tests for slope differences and Student’s t-tests, assessed SM score changes before and after resim events.

Results:

The algorithm effectively removed immobilization devices, enabling accurate 3D surface contouring and SM score calculation. The permuted learning framework demonstrated high predictive performance (Table 1), achieving an AUC of 0.907 (90.7% prediction accuracy) and a positive recall of 88.14%. Of the 197 resim events, 174 were correctly predicted, with an average of 1.85 fractions earlier per patient. Only 3 patients who underwent resim were missed by the model.

Notable features included reductions in beam-specific deviation metrics, with priority given to previous fraction scores. Days from planning CT, while less prominent, contributed to the model's ability to capture changes in patient surface geometry, reflecting the clinical impact of tumor reduction on resim events.

Conclusion:

This study introduces a permuted learning approach with surface mapping for predicting resim decision support in head and neck IMPT. The high recall and AUC demonstrate potential to enhance clinical workflows by predicting resim needs earlier. Future work will focus on dosimetric validation.

Table 1: Head and Neck Resimulation Decision Prediction Metrics