SliceBridge: context-consistent repair of corrupted slice intervals in T1-weighted MRI

Structural MRI scans often suffer from localized slice corruption due to motion, hardware, or reconstruction errors. These artifacts can bias downstream morphometric analyses, yet discarding the entire scan is inefficient. The authors investigate whether a deep learning model can selectively restore these corrupted intervals by leveraging surrounding intact slices, and whether coordinating the reconstruction of multiple slices improves through-plane consistency compared to independent slice generation.

The authors propose SliceBridge, a framework based on rectified flow matching. Unlike standard image restoration that might treat each slice independently, SliceBridge uses interval-level coordination to ensure that reconstructed slices form a coherent anatomical progression. This is achieved through three mechanisms: using interval-correlated initial noise, sharing flow time across the entire interval during training, and synchronizing sampling during inference. The model is conditioned on the surrounding intact slices and their relative positional information, allowing it to reconstruct missing axial content while leaving the rest of the volume untouched.

SliceBridge demonstrated superior performance compared to linear interpolation, edge-guided adversarial learning (EG-GAN), and an uncoupled flow model. It achieved higher structural similarity (SSIM) and peak signal-to-noise ratio (PSNR) at all tested interval lengths. Crucially, the interval-level coordination significantly improved through-plane consistency, reducing internal slice-to-slice gradient errors by 32.9% to 41.3% compared to the uncoupled flow model. In downstream morphometric tasks, SliceBridge effectively reduced the median error in regional brain volume estimates from 1.95% in corrupted volumes to 1.05%, providing a robust tool for salvaging otherwise usable MRI data.