W-STFNet: A Wavelet Transform-Based Regularized Hybrid Recursive Spatiotemporal Fusion Registration Network.

Abstract

Deformable image registration (DIR) is a crucial technique in medical image analysis and is particularly important for 4D-CT-guided lung radiotherapy, where accurate spatiotemporal alignment and deformation plausibility are required for downstream tasks such as dose accumulation. However, many existing learning-based methods are limited in modeling global spatiotemporal dependencies and in preserving fine anatomical structures under large respiratory motion.    METHODS: To address these issues, this paper proposes a wavelet transform-based regularized hybrid recursive spatiotemporal fusion registration network (W-STFNet). The proposed method incorporates SwinLSTM to effectively capture global spatiotemporal dependencies. To achieve better semantic integration of features across scales and time steps, a multi-scale spatiotemporal attention fusion (MSTAF) module is proposed, which improves the network's robustness and stability. Additionally, we design a novel frequency-domain loss function based on Discrete Wavelet Transform (DWT), which optimizes fine-grained structural matching by aligning high-frequency sub-bands, effectively improving the accuracy of high-frequency detail registration. The method is optimized in an unsupervised, patient-specific one-shot setting without anatomical annotations or multi-patient pretraining.    RESULTS: Experiments on two public 4D-CT datasets (DIR-Lab and POPI-model) show that W-STFNet achieves competitive registration accuracy and stable performance across cases with varying deformation amplitudes. On DIR-Lab, W-STFNet attains a mean TRE of     <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mn>1.13</mn> <mo>±</mo> <mn>0.72</mn> <mtext>mm</mtext></mrow> </math> , and on POPI-model a mean TRE of     <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mn>0.87</mn> <mo>±</mo> <mn>0.56</mn> <mtext>mm</mtext></mrow> </math> , substantially reducing the initial misalignment. A two-sided paired Wilcoxon signed-rank test further supports that W-STFNet differs significantly from several learning-based baselines under the reported settings, although the absolute differences should be interpreted with respect to image resolution and annotation uncertainty.    CONCLUSION: W-STFNet provides an annotation-free, patient-specific one-shot registration framework that achieves robust and competitive performance for 4D-CT lung DIR, particularly in handling image registration scenarios involving large deformations and complex temporal dynamics.