IML-UNet: A brain-inspired spatiotemporal collaborative encoding method for medical image sequence registration

Deformable medical image registration is important for radiotherapy planning, respiratory motion analysis, and organ function assessment. In medical image sequences such as 4D CT, organ motion involves large displacements, local non-rigid deformations, and continuous temporal changes. Registration requires accurate spatial matching together with temporal consistency and deformation stability across the sequence. However, most existing methods either process spatial and temporal information in separate stages or treat 4D registration as independent pairwise tasks, limiting spatiotemporal interaction and long-range temporal dependency modeling. We therefore propose a spatiotemporal joint registration method that models multi-scale spatial information and temporal dependencies within a unified framework to improve accuracy, stability, and sequence consistency under large-displacement conditions.

We propose IML-UNet, a deep learning-based method centered on a new recurrent cell, Inception MLP-like LSTM (iMLSTM), inspired by spatiotemporal collaborative encoding in partially overlapping neuronal populations. The iMLSTM integrates parallel multi-scale feature extraction with long-range temporal modeling to form fused spatiotemporal representations at each network level and reduce hierarchical feature degradation. A Progressive Frequency Architecture (PFA) flexibly adjusts the contributions of high- and low-frequency features across network levels to strengthen representation capability. During training, an Adaptive Weighted Multi-Scale Structural Similarity Index Measure (AM-SSIM) dynamically optimizes multi-scale similarity assessment. We also design efficient sequence registration strategies to capture dynamic features of 4D sequences and support diverse registration scenarios.

On two publicly available 4D CT datasets, our method achieved TREs of 1.09 mm and 0.81 mm, respectively, demonstrating superior registration performance over prior methods.

The proposed method delivers stable and accurate registration for lung 4D CT image sequences, with its main advantage lying in hierarchical intra-level spatiotemporal coupling rather than in the simple stacking of spatial and temporal modules. The observed performance gains are achieved with comparable model size and inference time, which strengthens the practical relevance of the method. It provides a promising framework for modeling complex spatiotemporal deformations in this setting. However, the current evidence is limited to a per-case optimization setting, and broader cohort-level generalizability remains to be further validated.