Session Item

Knowledge-based (KB) plan prediction models were used to assess differences of plan performances between institutes in the context of right whole breast (RWB) irradiation using tangential-fields (TF). The aim of the current investigation was to assess major quantitative parameters that influence inter-Institute variability, possibly suggesting robust and efficient metrics to estimate KB-model transferability. 

Ten Institutions set RWB-TF KB models by using RapidPlan (Varian Inc., V.16.1 and earlier), following previously shared methods as same, number of patients (> 70), contouring (national guidelines for CTV/PTV/OARs), techniques (wedged or FiF) and outlier elimination criteria. Two patients of each center (in total n=20, external from the training sets) were used to perform cross validation test using a dedicated Eclipse research station (V.16.1): DVH’s prediction bands of OARs (heart, ipsilateral lung, contralateral lung, contralateral breast) predicted from the 10 models (normalized to 40Gy/15 fraction scheme) were exported and inter-institute variability was quantified by looking to SD of several DVH and dose/statistics parameters. Varian Model Analytics, software was used for seven models (only available for earlier V16.1 version) to quantify the distribution of anatomical (PTV/OARs volumes: Vol) and dosimetric (PTV V95%, D99%, Dmax, SD: PTV_dose) parameters referred to the original training data sets. Then, the Principal Component (PC) of GED (Geometry-based Expected Dose) was derived from a “test KB model” based on the 20 patients considered in the cross-validation; in the same patients, the portion of the right lung in field (d_C) along the isocentric axial slice was also quantified. The association between inter-institute variability of the predicted mean ipsilateral lung doses (Dmean) and Vol, PTV_dose and PC, PTV_dose and d_C was investigated

Dosimetry/anatomical parameters showed different distributions for the seven models in the original data sets. In particular, one center always showed no overlap between PTV and ipsilateral lung. For the remaining 6 models, there was a clear association between ipsilateral lung average Dmean value and the PTV median value of D99%, as shown in Figure 1 (R²=0.78), with average values of Dmean ranging between 4.7 and 6.1 Gy. PC and d_C analyses didn’t show correlation between individual geometrical/anatomical features and inter-institute SD of Dmean, as summarized in Figure 2: the mean value of SD was 13.2%. 

The different attitudes in handling PTV dose coverage in RWB-TF reflects into mild/moderate differences in sparing ipsilateral lung between Institutes. Inter-institute variability in lung dose prediction appears to have no relevant dependence on patient's anatomy. This result, in addition to the limited variability between models in predicting ipsilateral lung DVH, suggests promising potential applications in multi-institutional benchmark models. This study is supported by an AIRC grant (IG23150).