S1603
Physics - Autosegmentation
ESTRO 2026
Results: Figure1 shows higher DSC and recall scores for datasets from the training distribution (TD). ADC only models segmented and detected as accurate as ADC+T2 models. T2 models had the worst performance. MRRN models trained from either institution had similar results on the public datasets. Ensembling the MRRN models did not substantially improve metrics. NnUnet models had highest performance on TD, but resulted in variable accuracy on the public datasets.
Moritz Westermayer 1,2 , Nolwenn Delaby 3 , Alain Sottiaux 4 , Selma Kara Mohammed 5 , Marie-Claude Biston 1,2 , Catherine Khamphan 6 , Sandrine Huger 7 , Sophie Chiavassa 8 , Igor Bessieres 9 , Ilyas Achag 10 , Marine Stadler 11 , Sayhaya Farzam 12 , Corinne Millardet- Martin 13 , Thomas Lacornerie 14 , Cathy Fontbonne 15 , Jean Marc Fontbonne 16 , Ludovic Madec 17 , Romain Castanet 18 , Alexandre Coutte 19 , Julian Biau 20,21 , Yoann Pointreau 22 , Anne-Agathe Serre 23 , Joel Castelli 3 , Anaïs Barateau 5 , Aurélien Badey 6 1 Department of Radiotherapy, Centre Léon Bérard, 28 rue Laennec 69373 LYON Cedex 08, Lyon, France. 2 INSA-Lyon, Université Lyon 1, Villeurbanne, CREATIS, Equipe Tomoradio, CNRS UMR5220, Lyon, France. 3 Department of Medical Physics, Centre Eugène Marquis, Rennes, France. 4 Department of Radiotherapy, CHU de Charleroi, Charleroi, Belgium. 5 CLCC Eugène Marquis, University of Rennes, Rennes, France. 6 Department of Medical Physics, Institut du Cancer Avignon-Provence, Avignon, France. 7 Department of Radiotherapy, Institut de Cancérologie de Lorraine - Centre Alexis Vautrin, Vandoeuvre les Nancy, France. 8 Department of Medical Physics, Institut de cancérologie de l'Ouest (ICO) Centre René- Gauducheau, Saint-Herblain, France. 9 Department of Radiotherpy, Centre Georges-François Leclerc, Dijon, France. 10 Department of Radiotherapy, Centre Georges-François Leclerc, Dijon, France. 11 Department of Radiotherapy, IUCT Oncopole, Toulouse, France. 12 Department of Radiotherpy, IUCT Oncopole, Toulouse, France. 13 Department of Raditherapy, Centre Jean Perrin, Clermont-Ferrand Cadex 1, France. 14 Department of Medicla Phyisics, Centre Oscar Lambret, Lille, France. 15 ENSICAEN, CNRS/IN2P3, Université de Caen Normandie, Cean, France. 16 ENSICAEN, CNRS/IN2P3, Université de Caen Normandie, Caen, France. 17 Department of Radiotherapy, Centre de la Baie Avranches, Avranches, France. 18 Department of Radiotherapy, Clinique Sainte-Clotilde, Saint-Denis, France. 19 Department of Radiotherpay, CHU Amiens Picardie, Amiens, France. 20 Department of Radiotherapy, Centre Jean Perrin, Clermont-Ferrand Cedex 1, France. 21 INSERM U1240 IMoST, Université de Clermont Auvergne, Clermont- Ferrand, France. 22 Institut interrégional de cancérologie (ILC), Centre Jean-Bernar, Le Mans, France. 23 Department of Radiotherpay, Centre Léon Bérard, 28 rue Laennec 69373 LYON Cedex 08, Lyon, France Purpose/Objective: Deep learning (DL)–based autosegmentation (AS) solutions for head and neck (H&N) organs at risk (OARs) have been shown to outperform conventional methods. They are nowadays widely implemented in clinical practice, improving both patient treatment and radiotherapy workflow efficiency [1]. The study aims to
Conclusion: The comparable DSC scores for MRRN models of either institution for ADC and ADC+T2 sequences with high recall scores on OOD datasets indicates a high generalizability for MRRN models. The high variability in DSC scores for the nnUnet models and lower recall scores on OOD datasets indicate poor generalizability. References: 1. Simeth, Josiah, et al. "Deep learning ‐ based dominant index lesion segmentation for MR ‐ guided radiation therapy of prostate cancer." Medical physics 50.8 (2023): 4854-4870.2. Litjens, G., Debats, O., Barentsz, J., Karssemeijer, N., & Huisman, H. (2017). SPIE-AAPM PROSTATEx Challenge Data (Version 2) [dataset]. The Cancer Imaging Archive. https://doi.org/10.7937/K9TCIA.2017.MURS5C L3. Keno Bressem, Lisa Adams, & Günther Engel. (2022). Prostate158 - Training data [Data set]. I Computers in Biology and Medicine (Version 1, Vol. 148, s. 105817). Zenodo. https://doi.org/10.5281/zenodo.6481141 Keywords: Prostate, Target, Generalization
Digital Poster Highlight 4421
Multicentric comparative analysis of 8 commercial deep learning-based autocontouring solutions for the delineation of head and neck organs-at-risk
Made with FlippingBook - Share PDF online