S2246
Physics - Intra-fraction motion management and real-time adaptive radiotherapy
ESTRO 2026
15840 sequences (4 patients) with a further 3780 for validation and 11610 for testing. Test predictions were assessed via residual SSIM loss near the treatment volume, and correlation to Amsterdam Shroud breathing traces. Results: The CBAM-CNN was able to predict a 3D+t series of DVFs relative to reference CT, based on short series of kV projections (~1.5s and ~2o of gantry arc). Motion- DRRs were well matched to source kV images and deformation fields reproduced the expected breathing motion (figure 2)
Proffered Paper 3125
Volumetric motion estimation from a single kV projection: Towards real-time intrafraction motion compensation and adaption Marcus Tyyger 1 , Sean J Rooney 2 , Bashar Al-Qaisieh 2 , Michael G Nix 2 1 Scientific Computing, Leeds Teaching Hospitals NHS Trust, Leeds, United Kingdom. 2 Radiotherapy Physics, Leeds Teaching Hospitals NHS Trust, Leeds, United Kingdom Purpose/Objective: Intrafraction motion is a major confounder to accurate dose delivery in abdominal and thoracic sites. Respiratory motion measurement typically involves surrogates (SGRT, breathing traces) or limited imaging (e.g. kV planar or phase/amplitude binned 4DCBCT). MR-linacs potentially offer true-4D imaging, but have limited spatiotemporal resolution and availability.We leveraged deep-learning to estimate full 3D+t motion[1,2] for liver SBRT patients, based on kV projection images from a conventional IGRT linac, with 182 ms resolution throughout treatment delivery. Material/Methods: A recurrent CNN (figure 1), based on Zakeri et al.[3], and leveraging Channel-Based-Attention-Mechanism (CBAM)[4] was trained self-supervised to predict 3D deformation vector fields (DVFs) corresponding to instantaneous internal motion states, from short sequences of kV images (n=7, dt = ~1.5s). Reference 3DCT was deformed by a spatial transformer network (STN) and used to differentiably project Digitally Reconstructed Radiographs (DRRs), using the Siddons Jacob method.
Residual testing SSIM errors between kV and DRR in the region of the liver SBRT treatment were low (0.220–0.314), and comparable to training (0.189– 0.288). Predicted breathing traces showed excellent Pearson correlation with Amsterdam Shroud[5] (test: 0.756-0.933, train: 0.822-0.956), recovering variable breathing depth and duration, and even accurately reconstructing coughing (* in figure 2). Conclusion: We presented a self-supervised recurrent CBAM-CNN to estimate patient internal motion in 3D+t, using kV projections images. This technology, coupled with interleaved kV imaging during RT delivery opens the way towards full 4D motion estimation and adaption both offline and real-time online, using conventional kV IGRT platforms. References: [1] Yang B et. al. Deep Learning Applications in Motion
Structural dissimilarity (LSSIM) between motion-DRRs and input kV images was minimised, whilst encouraging correlation with Amsterdam Shroud[5] derived breathing traces extracted from input kV. Regularisation constraints Lsm, Ljac and Lbone were applied to ensure physically plausible DVF smoothness, tissue compression and bone rigidity.Images were obtained on Elekta Agility XVi linacs. Model training (PyTorch, NVIDIA T4), used
Made with FlippingBook - Share PDF online