Two-View CT Reconstruction Based on a Noise-Adaptive Regularization-Guided Conditional Diffusion Model

In industrial manufacturing, computed tomography (CT) is an effective method for defect inspection. However, it usually requires 360° scan, which would consume a long scanning time and thus limit its application for online detection. To achieve extremely fast CT, two-view CT reconstruction is proposed, where only two orthogonal projections are adopted to reconstruct 3-D CT volumes. To compensate for the lack of projections, most existing works rely on learning methods, such as the U-Net and the generative adversarial network (GAN), to provide sufficient prior knowledge based on datasets. Recently, some diffusion model-based works have been developed due to the model’s excellent ability to generate complex images. However, most of them are based on the unconditional framework, and the traditional total variation (TV) regularization may not be suitable here. In this article, we propose a noise-adaptive regularization-guided conditional diffusion model (CDM) for two-view CT reconstruction. To be specific, we adopt the filtered back-projection (FBP) results as the condition to guide the generation and embed a TV regularization-based optimization to ensure data consistency. Moreover, we particularly design a triaxial noise-adaptive TV (NATV) regularization term to adapt to the denoising process during generation and develop a lightweight neural network-based solution to accelerate the reconstruction speed. To validate the proposed method, we establish a real piston dataset with a variety of defects and demonstrate that our method achieves a more precise reconstruction of both overall shapes and tiny structures compared to existing methods.