Integrating Multi-scale Spatial and Temporal Dynamics for Deformable Medical Image Registration

In various clinical settings, deformable medical image registration plays a crucial role. The process is notably complex when addressing non-linear, substantial deformations between inhalation and exhalation phases, compounded by the temporal dynamics inherent in periodic movements, as observed in lung 4D CT imaging. This paper introduces a deep learning model composed of Inception-ResNet and ConvLSTM to optimize lung 4D CT image registration. The model combines the multi-scale spatial feature extraction capabilities of Inception-ResNet with the temporal sequence analysis prowess of ConvLSTM, integrating spatial deformations and temporal dynamics to significantly enhance the network’s learning capacity. Furthermore, the Inception-ResNet architecture is enhanced by using various quantities of lightweight $3 \times 3 \times 3$ convolutional kernels to replace larger ones, conserving memory without compromising feature extraction quality. Experiments on the DIR-Lab database yielded an average target registration error (TRE) of $1.19 \pm 0.75 \mathrm{~mm}$, demonstrating the clinical application potential of this method. Compared to other published methods, this approach offers superior registration performance.