Abstract
Medical image registration (MIR) is a fundamental task in clinical workflows, enabling the alignment of disparate scans into a common coordinate system. While deep learning models, particularly U-Net variants, have accelerated this process, single-decoder architectures often struggle with high-precision alignment of complex anatomical structures. This paper introduces Tetrahedron-Net, a novel architectural framework that utilizes a dual-decoder strategy to achieve coarse-to-fine refinement. By integrating a second decoder that interacts with both the encoder and the primary decoder, Tetrahedron-Net significantly enhances registration accuracy and transformation regularity across multiple benchmarks.
Introduction
Accurate medical image registration is essential for longitudinal studies, surgical planning, and multi-modal data fusion. Traditional iterative optimization methods, while accurate, are computationally expensive. Deep learning has shifted the paradigm toward fast, learning-based regression. However, standard U-Net architectures—consisting of a single encoder and decoder—may not fully exploit multi-scale feature interactions necessary for capturing intricate deformations.
The proposed Tetrahedron-Net addresses this by appending a second decoder to the standard U-Net backbone. This "Tetrahedron" structure facilitates a hierarchical refinement process, where the first decoder generates a coarse alignment that is subsequently polished by the second decoder using high-resolution features from the encoder.
Methodology
The core of Tetrahedron-Net is its tripartite structure: an Encoder (Enc), a first Decoder (Dec1), and a second Decoder (Dec2).
- Encoder (Enc): Extracts hierarchical features from the concatenated moving and fixed images.
- First Decoder (Dec1): Operates as a standard decoder with skip connections from the encoder to produce an initial, coarse registration field.
- Second Decoder (Dec2): This component receives inputs from both the Encoder and Dec1. This dual-input stream allows Dec2 to treat the output of Dec1 as a prior, refining it with the detailed spatial information preserved in the encoder’s skip connections.
The authors evaluated several internal configurations for Dec2, including UNet++, UNet3+, and DenseUNet. The DenseUNet variant, characterized by dense feature propagation, demonstrated superior stability and performance. This architecture is backbone-agnostic and can be integrated into existing state-of-the-art models such as VoxelMorph and TransMorph.
Experiments and Results
The framework was validated on three structural brain MRI datasets: LPBA40, IXI, and OASIS. Standard preprocessing, including skull stripping and affine alignment via FreeSurfer, was applied to all volumes.
Quantitative Performance
Tetrahedron-Net consistently outperformed baseline models. On the LPBA40 dataset, the DenseUNet variant of Tetrahedron-Net improved the Dice Similarity Coefficient (DSC) by 2.3% over VoxelMorph and 0.6% over TransMorph. On the IXI dataset, the model achieved a state-of-the-art DSC of 0.765.
Transformation Regularity
Beyond accuracy, Tetrahedron-Net produced more physically plausible deformations. Analysis of the Jacobian determinants revealed a significant reduction in non-positive values, indicating smoother transformations with fewer instances of unrealistic tissue folding or "tearing."
Conclusion
Tetrahedron-Net represents a significant advancement in medical image registration by demonstrating that a dual-decoder approach can effectively bridge the gap between coarse alignment and fine-grained anatomical precision. The architecture's ability to enhance diverse backbones—from convolutional to transformer-based—suggests broad applicability. Future research will explore the extension of the Tetrahedron design to other dense prediction tasks, such as medical image segmentation and 2D computer vision applications.
Summary updated