A Simple and Interpretable DVH Prediction Model: Integrating Distance-to-Target Histograms with Longitudinal Spatial Features.

OBJECTIVES: To develop a simple, interpretable dose-volume histogram prediction model for intensity-modulated radiation therapy by integrating distance-to-target histograms with longitudinal spatial features, and to evaluate its accuracy and generalizability. METHODS: Using prostate cancer cases, dose-volume histograms for the rectum and bladder (excluding overlaps with the planning target volume) were predicted via nonlinear regression based on slice-specific dose-to-distance relationships derived using a signed longitudinal distance metric. The model was validated on 70 internal cases and 10 external cases from a public database. Prediction accuracy was assessed using mean absolute error and mean dose differences. To assess clinical utility, five outlier cases per organ were replanned using model predictions as guidance. RESULTS: Internal validation showed average mean absolute errors of 3.33% and 3.16% for the rectum and bladder, respectively, with mean dose differences within 0.3 Gy. External validation confirmed robustness, yielding mean absolute errors of 3.45% and 2.12%. Replanning the outlier cases successfully reduced the mean dose by a median of 5.44 Gy for the rectum and 3.04 Gy for the bladder. These actual dosimetric gains correlated with the predicted room for improvement (Pearson's r = 0.85). CONCLUSIONS: The proposed model achieved high accuracy across independent cohorts. Its transparency and simplicity make it a reliable tool for routine quality assurance, identifying suboptimal plans, and guiding dosimetric optimization. ADVANCES IN KNOWLEDGE: Incorporating a signed longitudinal distance metric into distance-to-target histograms achieves accurate and generalizable dose predictions. This approach provides a transparent, highly interpretable alternative to "black-box" deep learning models.