Unsupervised pelvic CBCT-to-CT translation via cross-slice attention for 3D context modeling

Abstract

Accurate CBCT-to-CT translation is important for image-guided and adaptive radiotherapy in prostate cancer, where treatment-day anatomy can differ substantially from planning CT. Most unsupervised methods are based on 2D slice-wise translation and therefore underuse volumetric context, while fully 3D models are often computationally prohibitive for routine training and deployment. To address this gap, we propose cross-slice attention based medical image translation (CAMIT), an unsupervised framework that incorporates 3D contextual information without full-volume 3D convolution. CAMIT adopts a two-stage strategy: modality-specific autoencoder pretraining to obtain compact latent representations, followed by latent-space domain translation with a cross-slice attention module that models long-range inter-slice dependencies from randomly sampled slices. We evaluated CAMIT on an unpaired pelvic CBCT/CT dataset (40 training, 5 validation, and 10 testing cases) using both quantitative and qualitative analyses. CAMIT achieved a peak signal-to-noise ratio of 27.45 and an structural similarity index of 0.67, outperforming representative state-of-the-art unsupervised baselines; statistical testing further confirmed significant performance gains. These results indicate that CAMIT provides an effective trade-off between 2D efficiency and 3D contextual modeling, supporting treatment-day anatomical assessment and adaptive workflow guidance.