Unsupervised Pelvic CBCT-to-CT Translation with Cross-Slice Attention

An unsupervised CBCT-to-CT translation model using cross-slice attention achieved PSNR 27.45 and SSIM 0.67 on pelvic imaging, outperforming 2D slice-wise methods and providing 3D contextual modeling for adaptive radiotherapy.

• The model (CAMIT) uses a two-stage approach: autoencoder pretraining and latent-space translation with cross-slice attention.
• On 10 test cases, it achieved a peak signal-to-noise ratio (PSNR) of 27.45 and a structural similarity index (SSIM) of 0.67, significantly outperforming 2D slice-wise translation methods.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.