Self-cascade latent Schrödinger bridge for CT field-of-view extension

Abstract

Objective. Computed tomography (CT) field-of-view (FOV) truncation leaves peripheral anatomy entirely absent, introducing systematic errors in radiotherapy dose calculation and body-composition analysis. Existing methods either rely on unavailable raw projection data or suffer from over-smoothing and hallucination under severe truncation. Approach. We propose the self-cascade latent Schrödinger bridge (SCL-SB), combining a latent image-to-image Schrödinger bridge (I 2SB) with a self-cascade mechanism in a single shared-weight U-Net. I 2SB constructs a diffusion bridge from the truncated-image distribution to the full-FOV distribution, enabling high-quality reconstruction in ten denoising steps ( ≈29× faster than a 100-step denoising diffusion implicit model baseline). The self-cascade mechanism runs two sequential rounds sharing identical weights: the first produces an initial extrapolation; the second refines it using the first-round output as a hint. A frequency-aware residual mechanism amplifies gradient contributions at high-frequency inter-round residuals to stabilise cascade training and recover anatomically faithful detail.The truncation radius is sampled from a stratified distribution spanning mild to severe truncation during training, enabling a single model to cover the full severity spectrum without radius-specific fine-tuning. Main results. On a head-to-chest CT dataset spanning three truncation severities, SCL-SB, trained as a single model without radius-specific tuning, achieves the highest structural similarity index measure (SSIM) and a high-frequency ratio (HF-Ratio) closest to ideal among all compared methods at every severity, running nearly 9× faster than the latest vision transformer-based baseline. Ablation shows that cascade training alone already improves the first round’s output without running the second round at inference, an effect we term cascade implicit enhancement. Relative to the same architecture trained without cascade (I 2SB), the model recovers more high-frequency anatomical detail (HF-Ratio +45.3%) and achieves a higher SSIM by 0.076 ( 9.0%) at severe truncation, at no additional inference cost. Significance. SCL-SB addresses the posterior-mean regression that blurs high-frequency anatomical detail in ill-posed FOV extrapolation, produces spatially continuous Hounsfield unit reconstructions free of the patch-boundary discontinuities introduced by Vision-transformer-based architectures, and avoids the spectral hallucination characteristic of generative adversarial network-based methods. Although downstream clinical validation on larger, multi-centre cohorts remains necessary, these properties can potentially improve image guidance in clinical applications such as adaptive radiotherapy, body-composition assessment, and spine surgery.