We propose a physically consistent, reversible image augmentation approach for an unsupervised computed tomography (CT) reconstruction deep learning framework. The framework has the unique property of calculating the loss in the projection domain, so no image-domain ground truth is required. Even though previous evaluation proved feasibility of unsupervised deep learning for inverse problems like CT reconstruction, it also showed frail behavior after data alterations. To increase training diversity without violating projection-image fidelity, we apply reversible image augmentations that preserve non-zero image pixels, record the sampled parameters, and invert the transforms on the U-Net++ output. We describe implementation details, parameter sampling schemes and practical constraints for CT data, and discuss how the approach permits richer, physically consistent augmentation for improving performance in unsupervised reconstruction.