ESTRO 2024 - Abstract Book

S4545

Physics - Machine learning models and clinical applications

ESTRO 2024

Conclusion:

An NTCP model predicting the risk of acute toxicity after moderately hypofractionated breast RT was validated. The model underwent rigorous temporal validation in an external cohort, confirming its predictive accuracy even for patients treated with advanced techniques. Hypertension and lymph node dissection represent factors that potentially alter the early inflammatory response. Also, the validation indirectly corroborated the hypothesis that skin irradiation is of primary importance in generating acute effects, and V20 Gy can be used to define dose constraints and limit the radiation-induced effects on breast cancer patients.

Keywords: model validation, acute toxicity, dose constraints

References:

The study was supported by AIRC (Associazione Italiana per la Ricerca sul Cancro), IG25951 and IG23150 grants and by Fondazione Regionale per la Ricerca Biomedica, project nr. 110 - JTC PerPlanRT ERA PerMed, GA779282. D Raspanti of Tema Sinergie is gratefully acknowledged for support in the use of MIM and MIMassistant software.”

2518

Mini-Oral

CBCT-to-CT Deep-learning generation with uncertainty map predictions

Cédric Hémon 1 , Caroline Lafond 1,2 , Anaïs Barateau 1,2 , Blanche Texier 1 , Joël Castelli 1,2 , Renaud de Crevoisier 1,2 , Jean Claude Nunes 1

1 Univ Rennes, LTSI INSERM UMR 1099, Rennes, France. 2 Univ Rennes, Centre Eugène Marquis, Rennes, France

Purpose/Objective:

Recent advances in deep learning models have shown promise in the generation of synthetic CTs (sCTs) for external radiotherapy (RT) clinical applications due to their remarkable predictive accuracy and computational efficiency.

Generation of sCTs from CBCT or MRI plays a key role in tackling various challenges in adaptive RT (ART), for segmentation or daily dose calculation. However, synthetic generation in medical imaging remains subject to uncertainties. An essential aspect for clinical applications is the incorporation of an additional layer [1] for uncertainty prediction which can be categorized into two main types: