Abstract
Standard convolutional neural networks (CNNs) in medical image registration often fail to account for the diverse mechanical properties of different anatomical structures. We present SACB-Net, which utilizes a novel 3D Spatial-Awareness Convolution Block (SACB). By employing unsupervised clustering to identify distinct tissue regions and generating adaptive kernels for each, SACB-Net achieves state-of-the-art performance, particularly in complex abdominal registration tasks where large deformations are common.
Introduction: The Limitation of Vanilla Convolutions
Medical image registration aims to find the optimal deformation field to align a "moving" image to a "fixed" target. While deep learning has revolutionized this field, most models rely on standard convolutions that apply the same kernel weights across the entire spatial domain. This "translation invariance" is a drawback in medical imaging: a kernel optimized for rigid bone structures is ill-suited for highly deformable organs like the liver or lungs. Standard CNNs are forced to learn an "average" representation, leading to suboptimal alignment in regions with complex anatomical variability.
Methods: The Spatial-Awareness Convolution Block (SACB)
The core of our approach is the SACB, which breaks translation invariance by tailoring kernels to local spatial contexts. The process involves three key stages:
- Spatial Context Estimation: Input feature maps are "unfolded" into local patches and grouped into $N$ clusters using an unsupervised K-Means algorithm. This allows the network to distinguish between different anatomical regions without requiring manual segmentation labels.
- Adaptive Kernel Generation: For each cluster, a Multi-Layer Perceptron (MLP) transforms the cluster centroid into a unique set of convolution weights and biases.
- Region-Specific Convolution: These custom kernels are applied to their respective spatial regions, allowing the model to "pay different attention" to different tissues.
SACB-Net integrates these blocks into a multi-scale pyramid architecture. At each level, a similarity matching module (utilizing dot-product attention principles) calculates displacements, which are then refined from coarse to fine scales.
Experimental Results
We evaluated SACB-Net on three 3D datasets: IXI (Brain MRI), LPBA (Brain MRI), and a challenging Abdomen CT dataset. SACB-Net consistently outperformed state-of-the-art models, including VoxelMorph, TransMorph, and LapIRN.
- Brain Registration (IXI): Achieved a Dice score of 0.769, surpassing previous benchmarks while maintaining a lightweight parameter count (1.11M).
- Abdominal Registration: In this high-deformation scenario, SACB-Net reached a Dice score of 0.588, a significant improvement over the next best model, ModeT (0.550), and the popular TransMorph (0.444).
Beyond overlap metrics, SACB-Net produced more physically realistic deformation fields. It showed significantly lower rates of negative Jacobian determinants compared to baseline models, indicating fewer instances of "folding" or impossible tissue overlaps.
Conclusion
SACB-Net demonstrates that incorporating spatial awareness into convolutions significantly improves registration accuracy, especially for complex, multi-organ anatomy. By learning to treat different tissues uniquely through unsupervised clustering, the model provides a robust and flexible framework for medical image alignment. Future work will explore the integration of SACB modules into Transformer-based architectures and other medical imaging tasks.
Summary updated