CL-GAN: A progressive curriculum learning approach for bone CT super-resolution.

Cone-beam computed tomography (CBCT) provides rapid, low-dose imaging with large anatomical coverage but lacks the resolution required for detailed bone assessment in rheumatoid arthritis (RA). Enhancing CBCT to match high-resolution peripheral quantitative CT (HR-pQCT) is challenging due to differences in resolution, noise, and artifact profiles, causing standard super-resolution approaches to struggle in training with instability. We propose CL-GAN (curriculum learning generative adversarial network) to train a Cycle-Consistent GAN for CBCT super-resolution. Training is structured into four stages, beginning with paired synthetic mappings and gradually introducing unpaired real CBCT and HR-pQCT images from healthy and RA-affected joints. Each stage progressively increases task difficulty to bridge domain gaps stably. Performance was evaluated using image quality metrics, trabecular bone morphometry, and blinded review by an experienced reviewer. Repeated measures ANOVA and post-hoc paired t-tests assessed significance across stages. Progressive training led to consistent improvements in image quality and trabecular bone metric accuracy, with all gains statistically significant (p<0.001). In blinded review, two reviewers detected all RA erosions while one missed one erosion, and all reviewers reported difficulty distinguishing enhanced CBCT from HR-pQCT. An ablation study showed that skipping curriculum stages impaired image quality, highlighting the importance of gradual domain adaptation. This study demonstrates that curriculum learning enables stable and effective training of a super-resolution algorithm for CBCT images in the context of RA. By enhancing image quality and structural interpretability, the proposed framework increases the amount of relevant information that can be extracted from CBCT, supporting its utility in RA analysis and research.