Layer-wise Gaussian representation for biomedical Computed Tomography modeling

Computed Tomography (CT) scans are a critical tool in modern medical diagnostics, offering non-invasive insights into internal anatomical structures. Implicit Neural Representations (INRs) have become widely used in medical imaging due to their ability to model continuous signals compactly and with high fidelity. This thesis draws inspiration from the recent Gaussian Splatting (GS) technique, which represents volumetric data using adaptable Gaussian primitives, to propose an explicit neural representation for medical imaging tasks. The method introduces a novel layer-wise architecture that provides increased flexibility during initialization and training, as well as improved interpretability. The results demonstrate that the continuous representation generated by this method achieves quality exceeding state-of-the-art alternatives.