Tomographic image reconstruction from projections using generative AI

Since the 1970s, Computed Tomography (CT) has been a primary imaging technique, but concerns about patient exposure to radiation have grown. Although iterative reconstruction methods offer adequate image quality with reduced doses, their effectiveness remains limited. Recently, Deep Learning (DL) has emerged as a promising alternative for greater dose reduction and improved image quality. Generative Adversarial Networks (GANs) have shown the potential for creating images from real references, making them suitable for tomographic reconstruction. In this work, a method using Conditional GAN was developed to generate images directly from low-dose sinograms, bypassing traditional algebraic methods. Models were trained for physical phantoms, abdomen and thorax scans, and neuro studies, using data from the Low Dose CT (LDCT-and-Projection-data) database. The method demonstrated capability in generating low-noise images from low-dose sinograms, effectively reconstructing volumes in clinically relevant timeframes, even in emergencies, and the potential to reduce metal and photon deficit artefacts.