S1602
Physics - Autosegmentation
ESTRO 2026
Digital Poster 4364
Generalizibility of Multi-Center Trained Deep Learning Models for Prostate GTV Segmentation Ruben Bosschaert 1 , Josiah Simeth 2 , Rita Simoes 1 , Eduardo Pooch 1 , Atilla Simko 3 , Mar Fernandez Salamanca 1 , Ivo Schoots 1 , Neelam Tyagi 2 , Uulke A Van der Heide 1 , Christian Gustafsson 4 , Harini Veeraraghavan 2 , Tomas M Janssen 1 1 Department of Radiation Oncology, The Netherlands Cancer Institute, Amsterdam, Netherlands. 2 Department of Medical Physics, Memorial Slaon Kettering Cancer Centre, New York, USA. 3 Department of Disagnostics and Intervention, Umea, Umea, Sweden. 4 Department of Medical Radiation Physics, Skane, Lund, Sweden Purpose/Objective: Focal boosting the gross tumor volume (GTV) in prostate cancer radiotherapy improves outcomes versus homogeneous dose delivery but requires accurate delineation on multi parametric MRI (mpMRI), which is time-consuming and subject to inter-observer variability. Deep learning (DL) models have potential to reduce contouring times and improve consistency. However, generalizability of these models on out-of- distribution (OOD) data is limited. Methods to enhance generalizability are essential for multi-institutional clinical and research studies. Hence, two DL models, Multiple Resolution Residual Network (MRRN)[1] and nnUnet were evaluated to assess generalizability across patient variations including imaging acquisitions, lesion volumes, and demographics by testing on OOD datasets.This study evaluated performance and generalization of MRRN and nnUnet models on prostate cancer GTV detection and segmentation with Gleason Score (GS) ≥3+4by using
Figure 1: Workflow times for patients from CT to contours completed.
Figure 2 Workflow times for patients from CT to first fraction. Conclusion: Delays in starting radiotherapy can be associated with an increase in the risk of local recurrence.As a result, it is prudent that these delays should be as short as reasonably achievable.3 This study has shown that both the contouring workflow time and the overall CT to treatment interval can be reduced with MVision Contour+TM and dynamic scheduling of patient start dates post treatment planning. References: 1. Langmack KA, Alexander GG, Gardiner J, McKenna A, Shawcroft E. An audit of the impact of the introduction of a commercial artificial intelligence-driven auto- contouring tool into a radiotherapy department. Br J Radiol. 2023;96(1145):20230023. doi:10.1259/bjr.20230023.2. Malone C, Nicholson J, Ryan S, Thirion P, Woods R, McBride P, et al. Real world AI-driven segmentation: Efficiency gains and workflow challenges in radiotherapy. Radiother Oncol. 2025 Aug;209:110977.doi: 10.1016/j.radonc.2025.110977 3. Chen Z, King W, Pearcey R, Kerba M, Mackillop WJ. The relationship between waiting time for radiotherapy and clinical outcomes: a systematic review of the literature. Radiother Oncol. 2008;87(1):3-16. doi:10.1016/j.radonc.2007.11.016. Keywords: Auto-contouring, Automation, Efficient
T2w and ADC images. Material/Methods:
MRI datasets of patients with prostate cancer (pCA) from two institutions were used for training (INT1: 224, INT2: 185 cases). Model instances were tested on held out data from the two institutions (INT1:169, INT2: 43 cases) and two public external datasets (ProstateX: 63, Prostate158: 21 cases) [2][3]. All datasets consisted of co-registered T2w and ADC images (b-values: 0-800, 50-800, 200-800, 50-1000) with GTV labels (median sizes: 2.1, 0.6, 1.1, 1.3cc). Model instances were also created using only ADC, T2w to assess added value of the sequences.Institutional data was split for training- validation-test, stratified for scanner type and GS. Both model architectures were trained with standard preprocessing and postprocessing.Lesions with Dice similarity coefficient (DSC) > 0.1 were considered detected. Accuracy was evaluated using recall, DSC, and 95% Hausdorff distance (HD95). We present DSC scores as median and inter-quartile range (IQR).
Made with FlippingBook - Share PDF online