Computer-Based Techniques for Spine Analysis in MRI
The capability of magnetic resonance imaging (MRI) to offer precise analysis of soft tissues, customize imaging sequences to meet particular diagnostic needs, orient imaging planes with pertinent anatomical structures, and eliminate any risks associated with exposure to radiation distinguishes it as an advantageous technique for analyzing the vertebral column.
Over the past 10-15 years, there has been a surge in publications on computerized techniques for analyzing the spine in response to the growing interest in MRI.
The vertebral column is composed of interconnected vertebrae that are stacked in a column-like structure and separated by intervertebral discs. The spinal canal, a tubular structure that contains the spinal cord, which is also surrounded by a layer of cerebrospinal fluid, is housed within the vertebral column. The vertebral column and the spinal canal/cord are approached conceptually in distinct ways because of their differing characteristics.
Intensity Variation in MRI
A difficulty encountered in MRI is the absence of quantitative imaging measurements similar to the Hounsfield units used in X-ray computed tomography. Although achievable, obtaining quantitative MRIs is challenging, which is why they are not yet frequently employed in clinical practice.
As a result, information related to appearance, such as intensity ranges, cannot be transferred between various imaging sequences or settings. Techniques that heavily rely on absolute intensities may necessitate re-parameterization if used in a different configuration or may not be applicable at all.
Additionally, the intensities of identical tissues can vary spatially within a single image. Imperfections during image acquisition can result in intensity non-uniformity or inhomogeneity, particularly in the lower thoracic regions. Techniques intended for larger portions of the spine must address this spatial variation in the relationship between intensity and tissue.
Partial Volume Effect
The partial volume effect, a well-known phenomenon, occurs at the boundaries between tissues. The intensity observed at a voxel is a combination of the intensities of neighboring tissues, weighed by their volume contribution to the imaged voxel.
To achieve the desired imaging goal with minimal stress to the subject or maximum throughput to the device, the size of voxels must be carefully balanced against other influencing factors, as it is a significant factor in acquisition speed.
In order to improve efficiency even further, acquisition is often performed anisotropically by increasing voxel size in a specific direction. The selection of a particular slice orientation is often determined by the imaging objective, such as transverse slicing for the measurement of cross-sectional areas of the spinal cord.
Influence of Noise
When addressing noise, explicit preprocessing techniques such as Gaussian smoothing or anisotropic diffusion are employed. An alternative approach to address noise is to use appearance features that incorporate spatial neighborhoods, including Histograms of Oriented Gradients, Viola-Jones’ Haar-like features, and over-complete Haar wavelet transform coefficients for vertebra/disc localization, as well as the FCM framework for disc and spinal canal segmentation. Either way, it is assumed that noise has a mean of zero and possibly a fixed variance.
MRI noise follows a Rician distribution, which is different from the assumption of Gaussian distribution made in many image processing techniques. When the signal-to-noise ratio is large, the MRI noise can be accurately approximated by a Gaussian distribution with a fixed variance and zero mean.
However, as the signal-to-noise ratio approaches zero, the noise tends to follow a Rayleigh distribution. Hence, methods that heavily depend on absolute intensities, such as Viola-Jones’ Haar-like features, over-complete Haar wavelet transform coefficients, and FCM, may require reparameterization at best when the structure of interest appears hyperintense in one sequence and hypointense in another.
MRI-compatible metallic implants, such as those made of titanium, are increasingly common nowadays, making it not uncommon to observe MRIs of such implants in the spinal area. The degree of localized imaging artifacts, such as black spots, can be expected to vary depending on the size and mass of the implant. It is unrealistic to expect a fully automated treatment of these cases.
A computerized approach can at best identify a problematic case and utilize additional manual guidance, as fully automatic treatment of such cases is not feasible. When a subject is lying in a stable supine position, accidental body motion is not very likely to occur, and respiratory motion effects can be prevented through appropriate image acquisition techniques, such as breath-hold examinations or respiratory gating.
However, when the position is less stable, as in the case of an intervention, these claims may be questionable. It is probable that the increased complexity of the problem will require the incorporation of supplementary information through manual intervention.
Learning And Invariance
A crucial aspect of computerized spine analysis is the acquisition and integration of information through learning methods. The information that is learned can be categorized into three types:
- Pose (location, orientation, and size): Statistics present a significant challenge as they require reference to a coordinate system, making it necessary to re-learn them for new datasets. Nonetheless, these statistics are incredibly useful in simplifying tasks such as localization and segmentation, particularly in spine analysis where subjects are in a stable supine position.
- Geometry (shape and part relations): The relationship for geometry information is also intricate. The statistics of the shape and part relations’ geometry are formulated to be invariant to a certain subset of the pose aspects. To be more precise, invariances are typically designed to capture location and orientation, but may not necessarily account for variations in size. This could potentially pose a challenge when dealing with subjects of vastly different sizes. However, the geometry information can be reused regardless of the data set, device, imaging sequence, or settings.
- Appearance (image intensity): One’s ability to achieve invariance in appearance information strongly relies on the type of features extracted from the image intensities. It can achieve the independence of intensity range shift and re-scaling mostly through magnitude-normalized measures, whereas plain intensity learning does not provide such independence.
Some of this information directly challenges the invariances that are typically desired in computerized analysis, such as the independence of the coordinate system transformation or image intensity scale, which can vary with changes in imaging sequences, settings, and so on. One must constantly balance between the learned information and the intended invariance while performing computerized analysis.
Alternatives To Learning
There are various options for obtaining shape and appearance information, including general learning techniques and parametric models. Expert knowledge can be employed to specify admissible ranges and bounds for pose information and part relations.
Learning provides detailed information but comes at the cost of significant training effort. It is still unclear whether the level of detail in learned information has a significant advantage compared to information specified by experts.
Assessing the accuracy of manually created ground truth would be facilitated by measuring the inter- and intra-rater variability. However, it’s challenging to compare computerized approaches and judge what can be expected from them because of the lack of literature reports on this topic. The majority of reported literature does not include reproducibility experiments to test the effects of alterations in method parameters and manual interaction.
Automated evaluation routines, systematic or randomized, are recommended to test reproducibility, along with scan-rescan experiments, and comparing results on different aligned sequences. These experiments are essential to evaluate computerized approaches effectively.
In the coming years, we anticipate that there will be a wider range of imaging settings and sequences used in the analysis of spines with MRI. As a result, it is necessary to develop reliable concepts that do not depend solely on absolute intensities or learned appearance information.
To achieve this, it is necessary to further utilize general appearance properties of vertebrae, discs, the spinal canal, and cord. his involves utilizing the recurring appearance along the vertebral column and various symmetrical properties of vertebrae, discs, the spinal canal, and cord. A potential upcoming trend could involve the use of interventional MRI, where subjects may assume different positions.
The computerized methods designed for interventional MRI should include a major component of manual guidance to cover a wide range of possible subject positions and provide effective means of online correction. To be more specific, it is necessary to introduce information about the position of vertebrae and discs, as well as the path of the spinal canal or cord, before or during intervention.
In addition, future work could incorporate general geometric properties, such as the compactness and connectedness of vertebrae and discs, as well as the adjacency relations among vertebrae, discs, and the spinal canal, to complement or possibly replace other types of information.
If you are interested in knowing more about Computer-Based Techniques for Spine Analysis in MRI you have come to the right place!
I am Vedant Vaksha, Fellowship trained Spine, Sports and Arthroscopic Surgeon at Complete Orthopedics. I take care of patients with ailments of the neck, back, shoulder, knee, elbow and ankle. I personally approve this content and have written most of it myself.
Please take a look at my profile page and don't hesitate to come in and talk.