**Title:** AI-based dose calculation in medical physics
**Presenter**: Julien Bert
**Date**: 08.06.2026
Participants
Julien Bert (JB)
Wojciech Krzemień (WK)
Konrad Klimaszewski (KK)
Aleksander Ogonowski (AO)
Dominik Muszyński (DM)
Mikołaj Morozowski (MM)
Roman Shopa (RS)
Lech Raczyński (LR)
Michał Obara (MO)
Aurélien Coussat (AC)
Renata Mikołajczak (RM)
Łukasz Adamowski (LA)
Agnieszka Syntfel-Każuch (ASK)
## Discussion
**KK:** *p. 21* — Is the size of the error correlated with the size of the organs?
**JB:** Yes, it is correlated because of the size of the Monte Carlo (MC) simulation.
---
**KK:** *p. 16* — You condition on the .... Did you consider other conditioning approaches, for example using a linear feature field? (The question referred to the U-Net architecture.)
**JB:** We did not try that previously.
**KK:** The question is whether the conditioning is relatively important, or whether the relevant information is already included in the fast MC simulation training data.
**JB:** The conditioning is very important, because the output varies significantly depending on the beam angulation, position etc. And if this information were not provided to the network, it would produce some kind of average of what it has learned.
**WK:** Can you comment on the practical applications of deep learning (DL) and its implementation in clinical practice? There might be psychological effects, and perhaps legal aspects as well. What is the general attitude of the medical community toward these methods?
**JB:** These approaches are accepted and even requested by the medical community. There are also more demanding use cases, for example, treatment planning immediately before therapy or even "in-treatment" planning. Adaptive planning is becoming increasingly important, and AI-based methods are generally well accepted.
**WK:** Can you comment on the robustness of DL methods? One issue that you partly discussed is that they may sometimes fail to reproduce all detailed patterns. On the other hand, there is also the danger that DL methods may introduce artefacts. Are there any accepted standards that would ensure that such behaviour is not present?
**JB:** There are no general standards of that kind. However, for some applications there are commonly used metrics, such as the Gamma Passing Rate shown on p. 32. What we are aiming for is to provide uncertainty maps alongside the predictions of the DL model. This would, of course, help identify potentially unreliable predictions.
**WK:** What is your personal opinion about the generalization problem? For example, what happens when the model encounters a sample that is completely outside the distribution covered by the training set?
**JB:** It is possible to compensate to some extent for out-of-sample situations. Much depends on the application. For example, in the case of beam-related applications, there are some obvious constraints that help. However, I agree that, in general, this remains a significant problem. The idea of using specialized models for specific applications (e.g., thorax imaging) rather than a single model intended for all applications should make such situations less common.
There are minutes attached to this event.
Show them.
