by M. Stewart
Why is diagnosis of CSM so important?
Over the last few years, we have talked about diagnosis of CSM a lot, including how the diagnosis of CSM can be difficult, as it can resemble so many other conditions. However, as you all know, this is an important topic of conversation, because as a general rule: the earlier CSM is diagnosed, the earlier an operation can be performed, the more spinal cord function which can be preserved and the better the chance of recovery.
As you know, a diagnosis is based on a combination of the neurological examinations (testing reflexes, sensation and motor control) and MRI imaging (see Figure 1). This method of diagnosis is ok, but it isn’t perfect. At the moment imaging alone can’t tell us whether you have myelopathy or the extent of it – at the moment we have to wait for symptoms to appear. This may be because present imaging techniques are good at showing us changes around the spinal cord (e.g. where the compression is), but features of the spinal cord or within the spinal cord are inconsistent and the ‘health’ of the spinal cord is not assessable.
What is diffusion tensor imaging (DTI)?
DTI is a relatively new technique, still based on MRI imaging, but analysed in a different way. It is able to use the movement of water molecules to identify structures, and therefore show different types of information. This is relevant for the spinal cord as it allows the neural pathways within the spinal cord to be depicted. This technique has been around a while, and researchers have been looking at its use .
Like with any new information, researchers have been trying to understand whether this offers relevance to helping patients
What is machine learning and how can it help us read DTI scans?
Machine learning is a new process, where computers are used to spot patterns in data which are just not appreciable to the human eye. It’s a ground breaking technique which is being applied to all sorts of different processes; Google recently purchased DeepMind, a specialist in this area for £400m .
A group from China have been using this process to look at DTI imaging in CSM. They tell the patients which scans come from healthy patients and which come from CSM patients, and the computer searches for differences . Following this development they compared the performance of the computer to a group of spinal surgeons, and found the computer was correct 93% of the time.
This was a very exciting result, especially as Wang et al didn’t tell the computer what specific information to use when figuring out if a scan showed CSM or not. They just used an average of information from across the whole spinal cord, without attempting to select those features of the DTI scan which are key for CSM. This is an issue, as if the computer is considering irrelevant aspects of the scan then this is likely to distort the actually relevant data and potentially lead to a misdiagnosis.
This year, Wang et al improved on their efforts  by figuring out what information the computer should pay attention to and what it should ignore. They used essentially same method as before:
1 It’s better to look at some parts of the cord, rather than the entire surface
In 2015 Wang et al’s computer looked at the entire surface of the cord (figure 4) and calculated the mean of what it saw. This time around, the divided the cord up into chunks called ‘voxels’ . They then used a computer algorithm to determine which voxels had the most power to distinguish healthy cord from CSM. The question is now how many voxels to look at – too many and the data all blurs together, too few and important points may be missed entirely. They found that computers spot CSM most accurately when they look at the 60 most important voxels of the cord.
Figure 5 – Earlier work told the computer to look at the entire cord to spot CSM (shown left). The results of this latest experiment suggest it would be most effective to look at just 60 voxels worth of cord, in the lateral portion of the cord’s white matter (shown right). Red lines show the area the computer is set to analyse. Image adapted from teachmeanatomy.info.
2 The edges of the cord are the key for spotting CSM
Wang et al divided the white matter of the cord up in lateral (edges), ventral (front) and dorsal (back) portions (figure 6), and told the machine to look at voxels in one of these three locations. They found that the computer spotted CSM best when it looked at the lateral portions. This agreed with another study .
3 Computers can spot CSM most accurately just by when they look at just one aspect of the DTI
MR scans and DTI scans actually work by looking at the movement of water. Information in DTI scans can actually be split up into types, depending on what aspect of water movement is being looked at. This study found that something known as ‘fractional anisotropy’ (basically a measure of how likely water is to move along a line in one direction) was the most important for distinguishing CSM spinal cord from healthy cord.
Altogether, these refinements made the computer 96% as accurate in diagnosing CSM as the group of senior spinal surgeons.
What does this mean for CSM patients?
Right now, not a huge amount. While the computer achieved some astounding results, neurological exams by senior surgeons still remain the gold standard for overall diagnosis of CSM. The researchers themselves stress that their work is only a “blueprint”. However, the results are promising as they suggest that DTI could eventually make its way into every hospital, without the need for expert interpreters.
Progress for machine learning will also greatly enhanced once we better understand CSM. For currently the computers are detecting patients based on our current diagnostic practice, which we know is limited; for some patients may not have compression on MRI imaging, or very subtle symptoms. But clearly this technique holds great promise for the future.
 Wang, S.-Q. et al. Prediction of myelopathic level in cervical spondylotic myelopathy using diffusion tensor imaging. J. Magn. Reson. Imaging 41, 1682–1688 (2015).
 Facon, D. et al. MR diffusion tensor imaging and fiber tracking in spinal cord compression. AJNR. Am. J. Neuroradiol. 26, 1587–94
 Wang, S., Hu, Y., Shen, Y. & Li, H. Classification of Diffusion Tensor Metrics for the Diagnosis of a Myelopathic Cord Using Machine Learning. Int. J. Neural Syst. 28, 1750036 (2018).
 Vedantam, A. et al. Diffusion Tensor Imaging of the Spinal Cord. Neurosurgery 74, 1–8 (2014).
Stay tuned for more informative and (hopefully!) exciting blogs next week!!!
The latest expert and patient articles