This invention is a novel deep-learning algorithm for computed tomography (CT) image reconstruction. The neural-field technique has a small graphic processing memory (GPU memory) footprint, which allows it to be applied in three-dimensional or higher-dimensional CT imaging. The proposed technique is self-supervised and trained instance-wisely, eliminating data labeling or any training of datasets. This algorithm can also produce synthesized projections that can be combined with existing projections to yield denser or complete projection data to be used by other image-reconstruction methods. Background: Each year there are more than 70 million CT scans performed in the U.S., with that number increasing annually at a rate of 10%. Despite several image-reconstruction algorithms for CTs existing, there is a lack of commercial products utilizing those techniques. This invention has the potential to change the standardized products by enabling faster imaging which, in turn, allows for more efficient diagnoses. The technology can be commercialized to enhance and improve cone-beam CT image reconstruction, which translates to a further reduction of radiation doses. Radiation from CT scans may cause damage to the DNA in one’s cells and thus increase their chances of these cells developing into cancer. Reducing radiation doses lowers the chances of negative effects from these scans. In addition to this, the per-data self-supervision component of the technique opens the door to its wide adoption across different imaging geometries, protocols, or scanners. Applications:
Advantages: