A HYBRID ALGORITHM OF FBP AND GENERATIVE ADVERSARIAL NETWORKS (GAN) TO COMPENSATE FOR DATA INSUFFICIENCY IN SPARSE-VIEW TOMOGRAPHY

This article proposes a novel hybrid algorithm for image reconstruction in computed tomography (CT) using sparse-view projection data. It is well known that the traditional Filtered Back Projection (FBP) method introduces severe streak artifacts in limited-angle sinograms. On the other hand, fully iterative Total Variation (TV) methods provide high-quality results but are highly computationally time-consuming. Taking this into account, we have developed a method that initially performs rapid base image recovery using FBP and subsequently refines it through a supervised Conditional GAN (cGAN) architecture. This FBP-GAN approach effectively eliminates noise and residual artifacts present in the base image. Conducted visual and quantitative analyses have demonstrated that the proposed method is significantly superior to the conventional U-Net and pure iterative methods. Specifically, the hybrid method preserves tissue-specific textures and fine details with high precision. Concurrently, the trade-off graph between reconstruction time and PSNR (Peak Signal-to-Noise Ratio) metrics confirmed that our method drastically saves computational time. The results indicate that the FBP-GAN algorithm holds great promise for reducing radiation doses and improving computational efficiency in medical diagnostics. In the future, it is planned to test this model under various sensor noise conditions encountered in clinical practice.