Spatial Normalization of Spine MR Images for Statistical Correlation of Lesions with Clinical Symptoms

Christos Davatzikos, Dengfeng Liu, Dinggang Shen, Edward H. HerskovitsRadiology 2002 · DOI

TL;DR

This paper introduces a novel method for spatially normalizing spine MRI scans, aligning them into a common "stereotactic space." This is crucial for accurately correlating spinal lesions with clinical symptoms by reducing anatomical variability between patients. The method uses a deformable spine model and elastic warping to achieve robust registration.

What Problem Does This Paper Address?

Low back pain is a widespread issue, yet its causes are often unclear, and definitive diagnosis is challenging. While advanced imaging techniques like MRI can reveal structural abnormalities, directly linking these findings to a patient's pain or symptoms is difficult. This is largely due to the significant anatomical variation among individuals, making it hard to compare lesions across different patients. To address this, researchers often use a "stereotactic space" – a standardized coordinate system – to analyze imaging data from large groups. However, existing methods for creating such a space for spine MRIs are limited. This paper aims to fill that gap by developing a robust spatial normalization technique for spine MRIs, enabling more accurate statistical analysis of lesion-symptom relationships.

Methods

The core of this method is an Adaptive Focus Deformable Model (AFDM). Here's a breakdown:

  • Deformable Spine Model Construction: A detailed 3D surface model of the spine (specifically vertebral bodies L1-S1 and the spinal canal) is created from a manually segmented MR image. This model serves as the "template" in the stereotactic space.
  • Statistical Training: The model is then "trained" on a dataset of normal spine MRIs. This process captures the statistical range of shape variations across individuals. Crucially, the model uses "attribute vectors" at each vertex to characterize local and global geometric properties, helping it avoid anatomically unrealistic deformations by seeking similar attribute patterns in new images.
  • Spatial Normalization via 3D Elastic Warping: Once trained, the AFDM is used to normalize new patient MRIs. The model is superimposed on a patient's image and deforms to match its edges. This deformation process is guided by the model's learned shape statistics and attribute vectors. The resulting transformation is then inverted to warp the patient's image into the template's stereotactic space. The paper notes that while the model itself is warped to the individual, the inverse transformation is used to map the individual to the template. Elastic warping is employed for interpolation to ensure smoothness and preserve relative structure locations.

Datasets

The paper describes using a dataset of 14 standard spoiled gradient recalled echo (SPGR) spine MR images from normal male volunteers (ages 25-37) acquired on a GE Signa 1.5T scanner. One image was used to create the initial template model, and the remaining 13 were used for training and validation. No new benchmark datasets were introduced.

Key Insights and Ideas

  • Attribute Vectors for Robustness: The use of attribute vectors is a key innovation. By encoding local and global geometric properties, these vectors help the deformable model avoid being "fooled" by spurious edges, leading to more anatomically plausible segmentations and registrations.
  • Hierarchical Deformation: The model's deformation process is hierarchical, with the spinal canal driving initial adjustments and the more complex vertebral bodies following. This approach has been shown to improve robustness against confusing image features.
  • Stereotactic Space for Spine Analysis: The paper highlights the critical need for a standardized stereotactic space for spine imaging, analogous to its established use in brain imaging, to enable meaningful statistical comparisons of lesions and symptoms.

Key Results

  • High Registration Accuracy: The method achieved an average overlap of 81.5% (with a standard deviation of 3.6%) between algorithm-defined and manually-defined regions for vertebral bodies and spinal canal segmentation.
  • Strong Correlation: The correlation coefficient between automatic and manual volume measurements was 0.98.
  • Minimal Volume Difference: The average difference between manual and algorithmic volume estimations was 11.7%.
  • Sharp Averaged Images: The average of the spatially normalized images showed high anatomical definition, indicating minimal residual inter-individual variability and accurate registration.
  • Tight Lesion Distribution: Simulated lesions placed in the L4-L5 disc showed a very tight spatial distribution after normalization, which is crucial for statistical parametric mapping methods that link lesion location to symptoms.

Limitations and Future Work

  • Limited to Gross Morphology: The current approach primarily accounts for the gross morphology of the spine. Future work could incorporate more subtle anatomical features.
  • Potential Impact of Lesions: While the algorithm is believed to be robust to lesions, its accuracy might be reduced when applied to patient data with significant lesions. Further validation on such data is needed.
  • Inclusion of Posterior Elements: The current model does not include posterior elements of the spine, which could be added to improve accuracy for structures like neural foramina and facet joints.
  • Validation on Patient Data: The paper suggests the need for further application and validation on patient data with significant lesions.
  • Statistical Analysis: The paper focuses solely on spatial normalization and does not detail the subsequent statistical analysis for lesion-symptom correlation. Future work could explore more elaborate statistical models.

Summary updated

Abstract

An image analysis method was developed for spatial normalization of spine magnetic resonance images. A deformable shape model of the spine is first constructed, and it is subsequently used by an automated algorithm to find a shape transformation that places patient data into a stereotactic space. Very good agreement with manual segmentations was observed. The main application of this method is in lesion-deficit analysis for determining associations between structural damage and clinical symptoms.
BibTeX
@inproceedings{davatzikos2002spatial,
  author = {Christos Davatzikos and Dengfeng Liu and Dinggang Shen and Edward H. Herskovits},
  title = {{Spatial Normalization of Spine MR Images for Statistical Correlation of Lesions with Clinical Symptoms}},
  booktitle = {Radiology},
  year = {2002},
  doi = {10.1148/radiol.2243011266},
}
Figure