S2068
Physics - Image acquisition and processing
ESTRO 2026
Digital Poster 2744
consistently achieved superior geometric alignment with CBCTs (NMI: 0.29-0.31, CC: 0.96-0.98 across test sets) compared to real-data trained models (NMI: 0.22- 0.28, CC: 0.87-0.97). Notably, pixel-wise metrics (MAE, PSNR, SSIM) favored models trained and tested on data with matching registration methods, meaning models trained on rigidly aligned data showed best results on the rigidly aligned test dataset (Table 1). Models trained on real data learned to reproduce registration errors, artificially improving pixel metrics while degrading geometric accuracy (Figure 1).
AI based MRI reconstruction in Radiotherapy: Improving Efficiency Without Compromising Quality Richard Speight 1 , Neil Sorby 2 , Helen Shepherd 2 , David Bird 1 , Sarah Wright 1 , David Higgins 3 , Bashar Al-Qaisieh 1 1 Medical Physics and Engineering, Leeds Cancer Centre, Leeds, United Kingdom. 2 Radiotherapy, Leeds Cancer Centre, Leeds, United Kingdom. 3 Clinical Science, Philips, Farnborough, United Kingdom Purpose/Objective: Magnetic resonance imaging (MRI) is increasingly used for radiotherapy planning due to its superior soft tissue contrast compared with CT, enabling more accurate target delineation and organ-at-risk contouring. However, conventional MRI sequences often require lengthy acquisition times, which can reduce patient comfort, increase motion artefacts, and limit departmental throughput. SmartSpeed AI, an AI- enhanced compressed sensing technique (AI CS- SENSE), offers to accelerate MRI acquisition while maintaining or improving image quality, potentially enhancing clinical workflow and patient experience. This study aimed to optimise, commission, and clinically validate SmartSpeed from Philips for MRI protocols across multiple anatomical sites, evaluating scan time reductions, image quality, geometric accuracy, and suitability for radiotherapy contouring. Material/Methods: Existing radiotherapy MRI protocols from a single centre were re-optimised by replacing conventional Compressed SENSE acceleration (CS-SENSE) with AI CS- SENSE. Optimisation strategies focused on reducing acquisition times while maintaining image quality and geometric fidelity. Additional site-specific objectives included improved spatial resolution for select anatomical sites and adaptation of prostate and liver protocols to support whole-pelvis coverage and breath-hold imaging, respectively. Candidate sequences were evaluated using phantom studies and in eight patients per anatomical site. Qualitative assessment was performed by radiographers, oncologists, radiologists, and a Medical Physics Expert (MPE). Quantitatively, geometric distortion was measured using the GRADE phantom, with a 2 mm tolerance within the clinically relevant field-of-view. Artefacts, including those associated with metal implants, were assessed, and clinical suitability for contouring was confirmed. Results: All previously commissioned sequences were successfully re-optimised with AI CS-SENSE. Scan time reductions ranged from 10–51%, with the greatest relative savings observed for anus (51%), brain (46%) and prostate (45%) imaging. An overview of all
Conclusion: Registration uncertainty in clinical training pairs forces networks to blur anatomically misaligned regions to minimize loss scores, compromising geometric fidelity. Synthetic data with perfect image alignment enables models to learn true anatomical transformations. Standard pixel-wise metrics inadvertently reward reproduction of registration errors, and thus mask this critical limitation. Geometric alignment metrics are
essential for evaluating the true CBCT-to-CT conversion quality in clinical applications. Keywords: registration uncertainty, CBCT-CT conversion
Made with FlippingBook - Share PDF online