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

Area of Science:

• Medical image analysis
• Deep learning for medical imaging
• Computational anatomy

Background:

• Deformable medical image registration is crucial for radiotherapy, respiratory motion analysis, and organ function assessment.
• 4D CT image sequences present challenges due to large displacements, non-rigid deformations, and temporal changes.
• Existing methods often struggle with spatiotemporal interaction and long-range temporal dependency modeling.

Purpose of the Study:

• To propose a spatiotemporal joint registration method for 4D CT image sequences.
• To improve accuracy, stability, and sequence consistency in registration under large-displacement conditions.
• To model multi-scale spatial information and temporal dependencies within a unified framework.

Main Methods:

• Developed IML-UNet, a deep learning method utilizing a novel Inception MLP-like LSTM (iMLSTM) recurrent cell.
• Integrated parallel multi-scale feature extraction with long-range temporal modeling for fused spatiotemporal representations.
• Employed a Progressive Frequency Architecture (PFA) and Adaptive Weighted Multi-Scale Structural Similarity Index Measure (AM-SSIM) for enhanced representation and training.

Main Results:

• Achieved superior registration performance on two public 4D CT datasets with TREs of 1.09 mm and 0.81 mm.
• Demonstrated enhanced accuracy and stability in registering lung 4D CT image sequences.
• Showcased comparable model size and inference time to prior methods.

Conclusions:

• The proposed method offers stable and accurate registration for lung 4D CT sequences through hierarchical intra-level spatiotemporal coupling.
• The approach provides a promising framework for modeling complex spatiotemporal deformations.
• Further validation is needed for broader cohort-level generalizability beyond per-case optimization.