ESTRO 2024 - Abstract Book

S3675

Physics - Dose prediction, optimisation and applications of photon and electron planning

ESTRO 2024

The model was successfully trained until convergence and allowed segment-wise dose modelling based on CT and irradiation segment data, as shown in Figure 1. Considering all 50 treatment plans of the test dataset, the model achieved median [range] gamma pass rates (GPR) of 98.9% [85.8%-100.0%], 93.5% [69.4%-99.7%] and 68.4% [40.92%-91.8%] for the 3%/3mm, 2%/2mm and 1%/1mm criterion respectively (Figure 2A-C). The best performance was achieved on prostate cancer treatment plans, exposing a median [range] GPR of 99.9% [95.1%- 100.0%], 98.9% [78.29%-99.7%] and 87.5% [48.3%-91.8%]. The model showed the least agreement with the MC simulation for the partial breast irradiation treatment plans, reaching median [range] GPR of 98.2% [96.9%-98.8%], 89.7% [86.7%-91.6%] and 59.3% [55.6%-70.0%]. Analyzing the dose distributions arising from single irradiation segments, we achieved median GPRs of 96.1% [51.2%-100.0%], 85.2% [32.9%-99.4%], 55.5% [13.8%-89.5%] on the whole test dataset (Figure 2D-E). The highest GPRs were reached again in the prostate dataset, where the model achieved median [range] GPRs of 99.2% [68.0%-100.0%], 93.8% [63.0%-99.4%] and 65.0% [19.4-89.5%]. The model performed with the least accordance to the MC-simulation on the out-of-domain lymph node dataset. In this case, the median [range] GPRs were 93.3% [51.2%-100.0%], 80.3% [32.9%-99.4%] and 50.2% [13.8%-89.4%].

Figure 1: Monte-Carlo simulated dose distribution (Target), the model prediction (Prediction) and the gamma map for a single segment (top) and the super position of all segments of a treatment plan (bottom).