Reconstructing Knee CT Volumes from Biplanar X-Rays Via Self-Supervised Neural Field

X-ray Computed Tomography (CT) provides rich anatomical details that facilitate early and accurate diagnosis. In certain medical situations, such as surgical navigation, it is not feasible to meet the extensive multi-directional data acquisition requirements of CT, an extremely sparse data measurement is preferred. Recent research has shown that it is possible to reconstruct 3D volumetric data from sparse views using Implicit Neural Representation (INR). However, INR is not able to share knowledge across scenes, which limits its performance in CT reconstruction from biplane X-ray images. Our model employs an architecture that includes a positional image feature extraction encoder network to overcome the challenge of INR’s limitation in highly underdetermined inverse problems. Our model can accurately capture the geometric structure and fine-grained details of CT scans due to the continuity and arbitrariness of the continuous neural intensity field. Moreover, the model is capable of rapidly reconstructing knee CT scans of new samples using the trained model.