Yuyao Zhang

Self-supervised Coordinate Projection Networks (SCOPE) represent an emerging paradigm for solving highly underdetermined problems in tomographic image reconstruction. In this presentation, the presenter takes Sparse-View Computed Tomography (SVCT) as a representative application to introduce the core methodology and innovations behind this framework. SCOPE leverages implicit neural representations to model continuous imaging signals from sparse, discrete measurements, effectively acting as an implicit regularization prior for the ill-posed inverse problem. By integrating the CT forward imaging model, SCOPE directly optimizes over spatial coordinates to reconstruct high-fidelity images from severely limited projection data. This approach not only eliminates the need for paired training data but also significantly improves reconstruction robustness across different scan geometries. Recent advancements have extended this framework to a range of medical imaging tasks, including accelerated Magnetic Resonance Imaging and metal artifact reduction in CT, achieving state-of-the-art performance. In this talk, the presenter will further explore key technical components that contribute to SCOPE’s success, such as the re-projection strategy for solution space enhancement, the role of coordinate encoding schemes, and efficient ray tracing implementations. Together, these innovations position SCOPE as a powerful, generalizable tool for next-generation low-dose medical imaging.