RA-BiMENet: Continuous-Time 4D Medical Image Interpolation via Relation-Aware Bi-Directional Motion Estimation.

Four-dimensional medical images introduce the temporal dimension to three-dimensional spatial data, enabling the dynamic characterization of organ motion and providing important support for disease diagnosis and functional assessment. However, due to constraints such as low-dose acquisition and prolonged scanning, the challenges faced in obtaining 4D medical images with high temporal resolution include insufficient spatial sampling, severe motion artifacts, and image blurring. Therefore, generating high-quality and temporally continuous intermediate frames while ensuring patient safety remains a critical challenge in 4D medical image interpolation. To address this issue, we propose the Relation-Aware Bi-directional Motion Estimation Network (RA-BiMENet) for 4D medical image interpolation, which enables the accurate prediction of intermediate frames at arbitrary time points. Specifically, RA-BiMENet consists of two key components: a spatiotemporal transform MLP (TS-MLP) module and a hierarchical spatiotemporal fusion (HSTF) module. The TS-MLP module performs bi-directional motion estimation in a pyramid-recursive manner, where a relation-aware multi-scale MLP (RAM-MLP) unit is introduced to model local correlations and multi-scale dependencies for accurate nonlinear motion estimation. Based on the estimated transformations, the HSTF module hierarchically integrates cross-temporal features through forward warping and self-attention, thereby enhancing local detail restoration while preserving global temporal consistency. Experimental results demonstrate that RA-BiMENet outperforms state-of-the-art methods on multiple quantitative evaluation metrics and is capable of generating high-fidelity and temporally coherent interpolated frames under complex deformation scenarios, validating its effectiveness and superiority for continuous-time 4D medical image interpolation.