S1589
Physics - Autosegmentation
ESTRO 2026
retrospective test cases to gauge potential utility.A prospective study was conducted to evaluate clinical utility, quantify time savings and perceived differences when consultant oncologists worked with or without AI contour starting points. The midpoint of the prospective study was used to trigger follow-up workflow commissioning, to ensure streamlined rollout if the study concluded that AI contouring was beneficial. Once implemented as routine practice, 6- monthly review procedures were put into place to continually monitor the performance of the solution in practice and trigger potential re-training/fine-tuning work.DSC scores are used to compare the unmodified AI auto contour to the clinically used contour.
Conclusion: AI contouring offers efficiency improvements for creating the CTV for craniospinal irradiation. Early data indicates that this solution, along with other automations, is reducing scan-to-treat times. The process used for this work is being used to develop whole ventricular irradiation target AutoContouring. We are currently evaluating further workflow optimisations which are enabled by target and organs- at-risk AutoContouring. References: [1] Isensee, F., Jaeger, P. F., Kohl, S. A., Petersen, J., & Maier-Hein, K. H. (2021). nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods, 18(2), 203-211. Keywords: AI,Craniospinal,AutoContouring Digital Poster 3818 Can input reconstruction predict segmentation error of deep learning models? Dany Rimez, John A Lee, Ana Maria Barragan Montero MIRO, UCLouvain, Bruxelles, Belgium Purpose/Objective: Uncertainty quantification (UQ) allows detecting wrong predictions of deep learning-based automatic segmentation, which is crucial for a safe clinical deployment.Most state-of-the-art methods - Monte Carlo dropout (MCDO), Deep Ensembles (DE), and Test-Time Augmentation (TTA) - rely on several predictions, implying high inference costs that may
Results: Retrospective study: The DSC score during training was 0.989±0.001 on the validation dataset (N=10) and 0.976±0.001 on the test dataset (N=18). Both consultant oncologists in the blinded qualitative review (N=18) gave 100% agreement that the AI contour would be “Quicker to fix than manual”. Prospective study: Clinicians took 91 minutes (mean, range: 51-165 min) to contour manually, with large variations between clinicians. Use of the AI contour resulted in a clinician-specific time decrease of 25±5%. All consultants (N=5) agreed in the prospective review that the AI contour would be “Quicker to fix than manual”. The model has been used in the treatment of 38 patients. Consultants were surveyed post- deployment to gauge satisfaction. Unanimous maximum positive scores were given for the solution evaluation and deployment questions. Post- deployment DSC and Surface DSC (Figure 2) are being monitored to ensure continued efficacy and clinician- specific tracking will be used to help identify potential issues around automation bias.
Made with FlippingBook - Share PDF online