Bi-Directional CT Image Transformation for Scalable and Realistic Data Generation

This work presents a bi-directional image transformation framework based on cycle-consistent generative adversarial networks (CycleGANs) for translating X-ray computed tomography (CT) data between simulated and real domains. The proposed approach addresses the persistent domain gap that limits the practical deployment of data-driven inspection methods in nondestructive testing (NDT). In the real-to-simulation direction, experimental CT data are transformed into a clean, simulation-like representation, enabling improved interpretability, precise annotation, and reliable algorithm validation under controlled conditions. Conversely, in the simulation-to-real direction, idealized CT volumes are augmented with realistic noise characteristics, contrast variations, and acquisition-related artifacts, facilitating scalable generation of high-fidelity synthetic datasets for robust model training. The framework integrates domain adaptation, artifact suppression, and intensity-preserving transformations to maintain geometric structure and material-related gray-value fidelity. Its effectiveness is demonstrated on both 2D radiographs and fully reconstructed 3D CT volumes using simulated and experimental aluminum specimens. Qualitative visualizations and quantitative gray-value profile analyses show strong agreement between ground truth and translated volumes, particularly at material boundaries and high-gradient regions. The proposed framework significantly improves data scalability, realism, and generalization performance, supporting the development of robust AI-driven CT inspection pipelines for industrial NDT applications.