S2309
Physics - Machine learning and AI algorithms
ESTRO 2026
Digital Poster 3611
Fast deformable vector field generation for lung ventilation imaging using deep learning and DEEDS registration. Gerardo U Perez Rojas 1 , René E Rodríguez 1 , Braian A Maldonado Luna 1 , Benito De Celis Alonso 1 , Jeremy S Bredfeldt 2 , Kamal Singhrao 2 1 Faculty of Mathematical Physics Sciences, Benemérita Universidad Autónoma de Puebla, Puebla, Mexico. 2 Department of Radiation Oncology, Brigham and Women's Hospital, Dana-Farber Cancer Institute, Harvard Medical School, Boston, USA Purpose/Objective: Accurate lung ventilation imaging is essential for functional assessment and adaptive radiotherapy in patients with respiratory motion. Image-intensity- based deformable image registration (DIR) methods such as Demons often require iterative optimization, resulting in long processing times and limiting real- time applicability [1,2]. DIR methods such as the Dense Displacement Sampling (DEEDS) DIR method are effective at accurately modeling lung deformation, exhibiting superior performance in capturing complex lung sliding motion compared to many other algorithms [3].To enable rapid DVF generation from 4DCT, we developed a hybrid deep learning model mediated with DEEDS DIR and benchmarked it against spline-based DIR methods. This approach aims to improve registration accuracy, reduce computation time, and support potential integration into time- sensitive clinical workflows such as adaptive radiotherapy and functional lung avoidance planning. Material/Methods: We developed a hybrid registration framework, DEEDS_DL, by using deformation vector fields (DVFs) generated from the DEEDS algorithm as the ground truth data to train a U-Net convolutional neural network (CNN). The resulting network learns to predict high-quality deformation grids, enabling fast and accurate registration for new images by leveraging the complex spatial correspondences originally captured by DEEDS.DEEDS_DL was validated in a 4D extended cardiac-torso (XCAT) anthropomorphic digital phantom and a database of 4DCT images of the patients with lung carcinoma [4,5]. The temporal variable was different phases of cardiopulmonary respiration (Figure 1). A model for medical image registration (DEEDS) was tested and compared with a hybrid model between deep learning and the DEEDS method of its own architecture for faster generation of DVF and Jacobian maps [1]. Model performance was quantified using image similarity metrics for deformed images (including Sum of Squared Differences (SSD), DVF generation time, and DVF generation accuracy using determinant of the DVF Jacobian matrix.
Results: The SSDs for the DEEDS_DL, DEEDs and Demons registration were 26.5 HU2, 26.1 HU2 and 30.6 HU2, respectively. All methods produced physically plausible deformations, with det(J) values near 1. The DVF generation time DEEDS_DL, DEEDS and Demons were 3 mins, 15 mins and 30 mins respectively.
Conclusion: Our hybrid framework successfully combines the robust deformation modeling of DEEDS with the inference speed of convolutional neural networks. The results demonstrate that the model produces anatomically plausible deformations with higher accuracy than standard DEEDS, while streamlining the registration process. Future work will focus on integrating this model into functional lung avoidance radiotherapy and enhancing cone beam CT image quality. References: [1] Balakrishnan, et al (2019). VoxelMorph: A learning framework for deformable medical image registration. IEEE Transactions on Medical Imaging, 38(8), 1788– 1800.[2] Vercauteren, et al. (2009). Diffeomorphic demons: efficient non-rigid registration. NeuroImage, 45(1 Suppl), S61–S72.[3] Heinrich, M. P., et al. (2013). MRF-Based Deformable Registration and Ventilation Estimation of Lung CT. IEEE Transactions on Medical Imaging, 32(7), 1239–1248.[4] Segars, W. P., et al. (2010). 4D XCAT phantom for multimodality imaging research. Medical Physics, 37(9), 4902–4915.[5] Hugo, G. D., et al. (2016). Data from 4D Lung Imaging of NSCLC Patients (Version 2) [Data set]. The Cancer
Made with FlippingBook - Share PDF online