NestedMorph: Enhancing Deformable Medical Image Registration with Nested Attention Mechanisms

TL;DR

NestedMorph addresses nonlinear intra-subject deformable registration between $T1$-weighted MRI and diffusion MRI by integrating high-resolution encoder details with decoder semantics via Nested Attention Fusion in a multi-scale CNN–Transformer framework. The method uses an encoder with Efficient and Channel Attention, a Dual Attention Transformer Block, a DAE-Former–LKA decoder, and a Spatial Transformer Network to estimate a deformation field $\boldsymbol{\phi}$ for warping the moving image toward the fixed image. Evaluated on the HCP Aging dataset, NestedMorph achieves state-of-the-art results with SSIM $=0.89$, HD95 $=2.5$, and SDlogJ $=0.22$, outperforming CNN-based, Transformer-based, and traditional methods. The work highlights the value of combining local and global feature processing with multi-scale fusion for accurate deformable registration and provides code for reproducibility.

Abstract

Deformable image registration is crucial for aligning medical images in a nonlinear fashion across different modalities, allowing for precise spatial correspondence between varying anatomical structures. This paper presents NestedMorph, a novel network utilizing a Nested Attention Fusion approach to improve intra-subject deformable registration between T1-weighted (T1w) MRI and diffusion MRI (dMRI) data. NestedMorph integrates high-resolution spatial details from an encoder with semantic information from a decoder using a multi-scale framework, enhancing both local and global feature extraction. Our model notably outperforms existing methods, including CNN-based approaches like VoxelMorph, MIDIR, and CycleMorph, as well as Transformer-based models such as TransMorph and ViT-V-Net, and traditional techniques like NiftyReg and SyN. Evaluations using the HCP dataset demonstrate that NestedMorph achieves superior performance across key metrics, including SSIM, HD95, and SDlogJ, with the highest SSIM of 0.89, the lowest HD95 of 2.5 and SDlogJ of 0.22. These results highlight NestedMorph's ability to capture both local and global image features effectively, leading to superior registration performance. The promising outcomes of this study underscore NestedMorph's potential to significantly advance deformable medical image registration, providing a robust framework for future research and clinical applications. The source code and our implementation are available at: https://github.com/AS-Lab/Marthi-et-al-2024-NestedMorph-Deformable-Medical-Image-Registration

Figures (4)

• Figure 1: Nested Morph Architecture. The NestedMorph architecture leverages dual attention mechanisms to fuse encoder and decoder features for accurate alignment between fixed and moving medical images. The overall architecture, shown here, highlights how Nested Attention Modules contribute to precise deformation field estimation.
• Figure 2: Nested Attention Fusion Module. The Nested Attention Module fuses local and global features extracted from both encoder and decoder outputs through a combination of spatial modulations and attention mechanisms. This design effectively captures both fine-grained local details and long-range dependencies, enhancing the deformable image registration process.
• Figure 3: Overall Framework of the Deformable Registration. This figure presents the complete NestedMorph registration pipeline, where the fixed and moving images are aligned using the NestedMorph network and a spatial transformer network to generate a refined deformation field. The architecture integrates both local and global information, improving registration accuracy for medical image modalities such as T1-weighted MRI and dMRI.
• Figure 4: Comparison of registration methods applied to medical images. The first two columns show the moving ($I_m$) and the fixed ($I_f$) image. Subsequent columns display registered images (top row) and their corresponding binary difference images (bottom row) for various methods. The SSIM values above the binary images quantify the similarity between the registered image and $I_f$. NestedMorph achieves the highest SSIM, indicating the best registration performance. Red boxes highlight key differences in the binary images.