ESTRO 2024 - Abstract Book

S3050

Physics - Autosegmentation

ESTRO 2024

Conclusion:

Our study addresses the everlasting challenge of target volume autosegmentation by harnessing histology content. Building on a deformable mapping approach, we transferred tumor contours, bridging the gap between these modalities, and providing our segmentation model with reliable labels. A standout feature of our work lies in the deployment of diffusion models and their capability to quantify uncertainty in segmentation. The predicted binary masks can be converted to probability maps, allowing for a nuanced characterization of tumor heterogeneity towards dose painting. Future works include clinical translation by moving to dose prediction and comparing it against traditional treatment plans, paving the way for new practices in H&N RT treatment planning. This work has benefited from a French government grant managed by the National Research Agency (ANR) and integrated into the France 2030 program with the reference ANR-21-RHUS-0005.

Keywords: Histopathology, Target Volume, Deep Learning

References:

[1] Leroy, A., Cafaro, A., Gessain, G., Champagnac, A., Grégoire, V., Deutsch, E., Lepetit, V. and Paragios, N., 2023, October. StructuRegNet: Structure-Guided Multimodal 2D-3D Registration. In International Conference on Medical Image Computing and Computer-Assisted Intervention (pp. 771-780). Cham: Springer Nature Switzerland. [2] Leroy, A., Cafaro, A., Champagnac, A., Classe, M., Gessain, G., Benzerdjeb, N., Gorphe, P., Zrounba, P., Lepetit, V., Paragios, N. and Deutsch, E., 2023. MO-0714 Statistical comparison between GTV and gold standard contour on AI based registered histopathology. Radiotherapy and Oncology, 182, pp.S584-S585.