Novel neural network‑based iterative image reconstruction method for low-dose CT

Computed tomography (CT) is a powerful imaging technique commonly used in clinical practice to examine the internal structures of the human body in a quantitative and non-invasive way. However, a significant concern surrounding CT imaging is the use of ionizing X-ray radiations, potentially harmful to patients. Even though CT scanner can deliver lower doses of X-rays during the acquisition process, traditional reconstruction algorithms present some limits when processing low-dose data, returning images of compromised quality and reduced clinical usability. As a result, there is growing interest in developing more advanced image reconstruction methods. Iterative algorithms and, more recently, artificial intelligence techniques have become key players in this area. Among them, deep learning neural networks result the most effective and commonly used approach for low-dose CT reconstruction. However, its performance heavily depends on the availability of large training data sets. This work introduces a novel iterative reconstruction method that combines the strengths of neural networks without requiring any training data. A phantom data was exploited to test the algorithm performance. It is composed of a main air-filled body and nine inserts of different materials. Results reported in this work prove the reconstruction ability of the algorithm and its effectiveness in low-dose scenarios. In this framework, it results particularly interesting in a clinical perspective to study particular acquisition geometry to preserve high-risk organs or operators. A tool to obtain the dose distribution inside the patient is also developed, in order to optimize the acquisition geometry.