Abstract
Determining whether input features are consistently strongly correlated with organ-at-risk (OAR) dose across different diseases can improve the accuracy of dose-volume histogram (DVH) prediction in knowledge-based treatment planning. In this study, we utilized patient geometric and dosimetric data to establish correlations with OAR doses and investigated the differences between distance-to-dose and dosimetric-to-dose relationships. Ninety nasopharyngeal cancer radiotherapy plans, 66 esophageal cancer radiotherapy plans and 52 rectal cancer radiotherapy plans were selected as study objects. Geometric feature was quantified by the distance-to-target histogram (DTH), and dosimetric feature was quantified by the conformal-plan-dose-volume histogram (CPDVH). The distance-to-dose correlation and dosimetric feature-to-dose correlation were calculated. DTH-DVH and CPDVH-DVH correlations were calculated for fractional OAR volumes of 30%, 50% and 60%. The correlations were calculated using the Pearson product-moment correlation coefficient (R). Compared to the distance-to-dose correlation, dosimetric feature achieved a stronger correlation (R) with OAR doses for 10 out of 13 OARs, including those to the brainstem (- 0.83, 0.93) and lung (- 0.60, 0.92). Compared to the DTH-DVH correlation, the CPDVH dose-volume achieved a stronger correlation (R) with the DVH dose-volume for 10 out of 13 OARs, including the brainstem (mean, - 0.67 vs. 0.80) and lung (mean, - 0.64 vs. 0.76). Compared to the CPDVH-DVH correlation, the DTH distance-volume showed a stronger correlation with the DVH dose-volume at the parotid (mean, - 0.88 vs. 0.60) and bladder (mean, - 0.56 vs. 0.46). The results indicated that the dosimetric feature showed a stronger correlation with most OAR doses than the distance-to-dose correlation, which is helpful for accurate DVH prediction. Patient geometry influences dosimetric feature-to-dose correlation when many OARs voxels are located inside the C-shaped target region or when the OARs significantly overlap with the target, and using only dosimetric feature for the DVH prediction of these OARs is not appropriate.