Search for a command to run...
Head and neck cancer (HNC) is one of the most challenging cancers to treat due to its complex anatomy and significant patient-specific changes during treatment. As the 6th most common cancer worldwide, HNC often has a poor prognosis due to late diagnosis and the lack of reliable predictive markers. Radiation therapy, typically combined with surgery, faces challenges such as inter-observer variability, complex treatment planning, and anatomical changes throughout the treatment process.Adaptive radiotherapy is essential to maintain precision as the patient's anatomy evolves during treatment. However, current low-invasive imaging methods before each treatment fraction, such as Cone Beam CT (CBCT) and biplanar X-rays, are limited in quality or provide only 2D images, making daily treatment adaptation challenging. This thesis introduces novel deep learning approaches to reconstruct accurate 3D CT images from biplanar X-rays, enabling adaptive radiotherapy that reduces radiation dose, shortens acquisition times, lowers costs, and improves treatment precision.Reconstructing 3D volumes from biplanar X-rays is inherently challenging due to the limited information provided by only two projections, leading to significant ambiguity in capturing internal structures. To address this, the thesis incorporates anatomical and deformation priors through deep learning, significantly improving reconstruction accuracy despite the very sparse measurements.The first method, X2Vision, is an unsupervised approach that uses generative models trained on head and neck CT scans to learn the distribution of head and neck anatomies. It optimizes latent vectors to generate 3D volumes that align with both biplanar X-rays and anatomical priors. By leveraging these priors and navigating the anatomical manifold, X2Vision dramatically reduces the ill-posed nature of the reconstruction problem, achieving accurate results even with just two projections.In radiotherapy, pre-treatment scans such as CT or MRI are typically available and are essential for improving reconstructions by accounting for anatomical changes over time. To make use of this data, we developed XSynthMorph, a method that integrates patient-specific features from pre-acquired planning CT scans. By combining anatomical and deformation priors, XSynthMorph adjusts for changes like weight loss, non-rigid deformations, or tumor regression. This approach enables more robust and personalized reconstructions, providing an unprecedented level of precision and detail in capturing 3D structures.We explored the clinical potential of X2Vision and XSynthMorph, with preliminary clinical evaluations demonstrating their effectiveness in patient positioning, structure retrieval, and dosimetry analysis, highlighting their promise for daily adaptive radiotherapy. To bring these methods closer to clinical reality, we developed an initial approach to integrate them into real-world biplanar X-ray systems used in radiotherapy.In conclusion, this thesis demonstrates the feasibility of adaptive radiotherapy using only biplanar X-rays. By combining generative models, deformation priors, and pre-acquired scans, we have shown that high-quality 3D reconstructions can be achieved with minimal radiation exposure. This work paves the way for daily adaptive radiotherapy, offering a low-invasive, cost-effective solution that enhances precision, reduces radiation exposure, and improves overall treatment efficiency.