ESTRO 2024 - Abstract Book

S4494

Physics - Machine learning models and clinical applications

ESTRO 2024

A multifractionated treatment is performed by accumulating doses from 30 consecutive validation episodes and Dose Surface Histogram (DSH) are generated for all target surfaces and critical structure (GTV, PTV, PTVs and OAR). The difference between the Desired DoseMap and the delivered dose is used to visualise over- and under-dosed areas.

Results:

Trained models demonstrate the following metrics results:

Figure 2 shows the overdosage of the PTV center while its extreme borders constitute the missing doses in the PTV DSH. The extra dose on the PTVs is only present on half of the surface and OAR receives a minimal dose. The movement and its uncertainty do not divert the beam from its objective, given the dose concentrated on the target areas.

Conclusion:

Given the spared OAR and the overdosed areas concentrated in the PTV, these first simplified simulations show interesting prospects. We suppose the dose homogeneity in the GTV and PTV to be limited by the small number of pixels in this synthetic target. The next step would be to extend the delivering problem for 3D doses and patient images.

Keywords: Reinforcement Learning, mobile tumors, real-time