ESTRO 2025 - Abstract Book

S77

Invited Speaker

ESTRO 2025

exceeding 8 Hz. This task demands high accuracy and robustness to accommodate large non-rigid motion and ensure the precise sparing of critical organs.

Current clinical methods for real-time tumor tracking rely primarily on deformable image registration (DIR) or template matching techniques to propagate tumor contours from an initial labeled frame to subsequent time resolved frames. While these methods are effective in certain cases, they can struggle with large-scale motion. In contrast, artificial intelligence (AI) methods that can perform rapid inferences during the treatment process hold great promise for overcoming these limitations [4]. By shifting the computational burden to the training phase, AI based approaches could strengthen real-time, accurate tumor segmentation, thereby improving the overall efficacy of MRI-guided radiotherapy. The primary goal of TrackRAD2025 is to facilitate the development and testing of algorithms capable of segmenting tumors in real-time from time-resolved sagittal 2D cine- MRI sequences by providing a comprehensive, multi institutional platform for evaluation. These algorithms will be given a template tumor segmentation on the first frame of the sequence and must be able to generate segmentations for subsequent frames without human intervention. The competition will provide both unlabeled and labeled cine-MRI datasets to support the development of machine learning models, with 477 patient cases included in the unlabeled dataset and 108 cases featuring manually segmented tumors. Data will be sourced from six international institutions, ensuring a diverse and representative sample of clinical cases. The TrackRAD2025 dataset will be divided into a public training set and a private test set. MRI data from both 0.35 T and 1.5 T MRI-linac systems will be included, covering a variety of tumor types and motion scenarios. The evaluation platform will incorporate a preliminary testing phase, where models will be assessed using sample cases, followed by a final testing phase with 50 distinct patient cases. Algorithm performance will be assessed based on several metrics, including the Dice similarity coefficient, Hausdorff distance, average surface distance, tumor center of mass, estimated radiation dose delivery, and inference speed. TrackRAD2025 represents the first public, multi-institutional dataset and evaluation platform for the competitive assessment of AI-based methods for real time tumor segmentation in cine-MRI. The outcomes of this initiative will provide critical insights into the most promising methods. Ultimately, the successful development of real-time tumor segmentation algorithms could enable more precise radiation delivery, supporting the replacement of beam gating with continuous tumor tracking via multi-leaf collimators. This advancement would reduce treatment times, increasing the potential number of treatments per day and improving overall patient outcomes.

Figure 1. The TrackRAD2025 Grand Challenge.