The analysis of micro-computed tomography (CT) scans for defect detection in complex materials such as fiber-reinforced polymers (FRPs) is challenging due to the need for pixel-accurate annotation and the inherent complexity of damage evolution. This work presents a novel, self-supervised methodology based on a cycle-consistent generative adversarial network (CycleGAN) to virtually transform fatigue damage states in FRPs within the high-cycle fatigue (HCF) and very-high-cycle fatigue (VHCF) regimes. The approach allows synthetic damage to be created or existing damage to be removed, facilitating defect identification by comparing high and low damage states. Without relying on manually labeled segmentation masks, damage regions are detected by isolating features removed during transformation, enabling efficient and objective defect segmentation directly from CT data.
The synthetic CT volumes generated are visually comparable to real scans, providing a robust basis for model training and material analysis. By applying multi-step transformations, different fatigue damage states can be emulated and differentiated, improving the understanding of damage initiation and progression. This work also shows that while some damage evolves chaotically across CT slices, other regions maintain semantic consistency, potentially reflecting underlying material behavior not captured by visible boundary conditions. Overall, this approach reduces the need for labor-intensive annotation, supports intuitive visualization of damage evolution, and enhances automated damage detection in FRPs. It paves the way for advanced applications in structural health monitoring, fatigue life prediction, and materials science through data-efficient, deep learning-based analysis of CT images.