Attention-based Shape-Deformation Networks for Artifact-Free Geometry Reconstruction of Lumbar Spine from MR Images

TL;DR

This work tackles automated, geometry-based reconstruction of lumbar spine anatomy from MR images, addressing segmentation-artifact issues that hinder parameter measurement. It introduces two attention-driven template-deformation networks, UNet-DeformSA and TransDeformer, and a TransDeformer-based error estimator to enable artifact-free meshes with mesh correspondence and quality control. A novel attention mechanism with relative-position embedding supports cross-modal interactions (image and contour tokens) through components ISA, SSA, and S2IA, with a three-stage training and a two-stage inference strategy. The approach yields accurate, artifact-free geometry outputs and enables reliable measurement of disc-degeneration parameters, with a practical QC tool to guide clinical use and potential extension to 3D data.

Abstract

Lumbar disc degeneration, a progressive structural wear and tear of lumbar intervertebral disc, is regarded as an essential role on low back pain, a significant global health concern. Automated lumbar spine geometry reconstruction from MR images will enable fast measurement of medical parameters to evaluate the lumbar status, in order to determine a suitable treatment. Existing image segmentation-based techniques often generate erroneous segments or unstructured point clouds, unsuitable for medical parameter measurement. In this work, we present $\textit{UNet-DeformSA}$ and $\textit{TransDeformer}$: novel attention-based deep neural networks that reconstruct the geometry of the lumbar spine with high spatial accuracy and mesh correspondence across patients, and we also present a variant of $\textit{TransDeformer}$ for error estimation. Specially, we devise new attention modules with a new attention formula, which integrate image features and tokenized contour features to predict the displacements of the points on a shape template without the need for image segmentation. The deformed template reveals the lumbar spine geometry in an image. Experiment results show that our networks generate artifact-free geometry outputs, and the variant of $\textit{TransDeformer}$ can predict the errors of a reconstructed geometry. Our code is available at https://github.com/linchenq/TransDeformer-Mesh.

Figures (10)

• Figure 1: Comparison of our geometry reconstruction approach (right) and the existing image segmentation approaches (left) for lumbar spine MR image analysis.
• Figure 2: Structure of UNet-DeformSA. The lumbar spine geometry is reconstructed gradually through three geometry-deformation modules. Each of the modules has shape self-attention layers.
• Figure 3: Structure of TransDeformer.
• Figure 4: The overview of the image self-attention (ISA) module.
• Figure 5: The overview of the shape self-attention (SSA) module