ESTRO 2025 - Abstract Book

S2485

Physics - Autosegmentation

ESTRO 2025

2631

Proffered Paper Real-time AI-based segmentation of prostate CTV on 2D-cine-MR Alexander Köhler 1 , Ivan Ćorić 1 , Nicole Wernlein 1 , Dominik Langner 1 , Simon Böke 2 , Cihan Gani 2 , Maximilian Niyazi 2 , Daniela Thorwarth 1,3 1 Section for Biomedical Physics, Department of Radiation Oncology, University Hospital Tübingen, Tübingen, Germany. 2 Department of Radiation Oncology, University Hospital Tübingen, Tübingen, Germany. 3 Cluster of Excellence “Machine Learning”, Eberhard Karls University Tübingen, Tübingen, Germany Purpose/Objective: Accurate delineation of the prostate and monitoring of its motion during radiation therapy are crucial for delivering ultra-hypo-fractionated treatment. MR-guided radiotherapy (MRgRT) offers the potential to address these challenges by providing real-time MRI and motion management. This work investigates the use of AI-based real time segmentation of the prostate on 2D-cine-MRI by comparing different model architectures and optimizing their inference speed. Material/Methods: The study utilized 2D-cine-MRI scans acquired during prostate MRgRT with a bTFE sequence and an acquisition frequency of 400 ms per frame from 43 patients. The training dataset consists of 100 images per orientation, obtained from 33 patients, with 40 manually contoured and 60 initially segmented by a nnUNetV2 trained on the 40 manual contours and subsequently manually corrected. Training dataset was split into an 80/20 train-validation split for model training. The test dataset consists of 30 manually contoured images per orientation from 10 unseen patients. Five model architectures were trained, the best being chosen for the inference time optimization: a standard nnUNetV2, a nnUNetV2-ResEnc, a MedNeXT and an UxLSTM were trained without modification of the network architectures using 500 epochs and five-fold-cross-validation. Models were trained for each architecture and cine MRI orientation and were evaluated on the test dataset using Dice score (DSC) and 95% Hausdorff (HD95%) distance. For subsequent optimization of the inference pipeline, the best performing model was streamlined for single-image prediction, enabling predictions directly from a single NumPy array and retaining model parameters across inferences. Stages and feature sizes were consecutively reduced to decrease computation time. A time measurement system tracks the key inference steps. 1000 single-image-inferences were performed on a Nvidia-RTX A6000 to assess optimization impact. Results: The best-performing model, nnUnetV2, achieved on the coronal/sagittal/transversal test dataset a mean DSC and HD95% of 94.30/83.38/92.05% and 2.65/5.36/3.80 mm, respectively. The other architectures (nnUNetV2-ResEnc, MedNeXT, UxLSTM) demonstrated comparable performance across all orientations, by achieving mean DSC ranging from 76.50% to 94.15% and mean HD95% values between 2.56 mm to 12.69 mm (figure 1). Using the optimized workflow reduced inference time from 6.0 s to 0.13 s per image (table 1).