ESTRO 2024 - Abstract Book

S3937

Physics - Image acquisition and processing

ESTRO 2024

modified to account for the fact that the REMs were simulated for a field of 26 x 26 cm 2 and 1 mm spot spacing, while the PR measurements were obtained for a 4 x 4 cm 2 field and 5 mm spot spacing.

Ray-tracing IDDs were simulated, delivering one spot every 5 mm for a 4 x 4 cm 2 field of 81 spots around the treatment isocenter. For each spot, the range error was computed, and the comparison was performed in terms of relative range error (RRE) with respect to the water-equivalent path length (WEPL) of each proton spot passing through the patient [1,2].

Results:

After 90 epochs, a low loss value was achieved during CNN training. There was no evidence of overfitting throughout the training. This indicates that the EfficientNet-B1 effectively processed the range error maps, correlating them with the assigned labels, and subsequently made accurate predictions. The performance metrics for multi-label image classification achieved 82% accuracy, 93% precision, 95% recall, and an F1-score of 93%. The excellent performance metrics demonstrated the AI-tool’s superior capability to detect multiple sources of proton range errors using PR images compared to other studies [1]. Furthermore, the proposed AI tool took 6 seconds to interpret a PR displaying its potential use in online adaptive workflow. Retrospective analysis of the PR measurements (Figure 2) showed that the simulated Ray-tracing IDDs for most patients were in agreement with measured IDD for spots located in anatomically stable areas. The agreement was mainly within a ±3% margin, which aligns with the range uncertainty threshold used for robust Monte Carlo-based optimization in our clinic.