Author: Xu Chen, Jun Lian, Yunkui Pang, Pew-Thian Yap 👨🔬
Affiliation: University of North Carolina at Chapel Hill, Huaqiao University 🌍
Purpose: Unsupervised CBCT-to-CT translation in the pelvic region is essential for accurate radiotherapy delivery and adaptive image-guided interventions. However, current models for cross-modality translation are predominantly constrained to 2D due to computational limitations, resulting in the loss of critical 3D spatial information. This study introduces a novel cross-slice attention mechanism designed to harness 3D volumetric data, significantly improving the accuracy and reliability of unsupervised CBCT-to-CT translation.
Methods: We developed Cross-slice Attention-based Medical Image Translation (CAMIT), a new unsupervised image synthesis method that leverages 3D information through an innovative cross-slice attention mechanism. Unlike traditional 2D methods, CAMIT incorporates information from multiple randomly sampled slices within the 3D CBCT volume, enabling the model to capture long-range inter-slice dependencies. This approach generates richer, 3D-aware representations, leading to substantial performance improvements in medical image translation. We validated CAMIT on a dataset of 55 paired pelvic CBCT and CT images from prostate cancer radiotherapy and compared it with eight state-of-the-art unsupervised image-to-image translation methods, namely CycleGAN, DRIT++, DCLGAN, Kang et al., UNIT, MUNIT, NICE-GAN, and SynDiff. Performance metrics included Mean Absolute Error (MAE, measured in HU), Peak Signal-to-Noise Ratio (PSNR), and Structural Similarity Index (SSIM).
Results: Our method demonstrated significant improvements over existing approaches, achieving an MAE of 43.97 ± 5.77 HU, PSNR of 27.45 ± 1.62, and SSIM of 0.67 ± 0.04. Statistical testing with a paired Student’s t-test confirmed the improvements were highly significant (p < 0.01) compared to all eight competing models.
Conclusion: CAMIT underscores the critical role of incorporating 3D information into unsupervised medical image translation. This innovative approach significantly enhances the accuracy of pelvic CBCT-to-CT translation, offering substantial benefits for improving treatment accuracy and adaptive replanning.