As we know Tractography is the only non-invasively tool for determining and measuring the pathways in the brain through observations of local diffusivity. The uncertainty of tractography which arises for example from noise.
I'd like to know what is this noise with example if possible.
Diffusion is a pretty noisy technique anyway (low signal to noise ratio). From the acquisition standpoint you have noise coming from the pulse sequence used to collect it (typically a single-shot EPI), which can have spikes or distortions caused by sinuses and phase encoding direction. Also the b-value itself can introduce noise, as the higher the b-value the lower the SNR for both GM and WM structures. We try to compensate for some of these by collecting multiple averages of each direction... but the noise is substantial, and that's just on the collection side of things.
When it comes to actually estimating the tensors, the metric you are using (FA, ADC) can also have issues depending on a variety of factors impacting your data.
Some useful papers:
Soares, J. M., Marques, P., Alves, V., & Sousa, N. (2013). A hitchhiker's guide to diffusion tensor imaging. Front Neurosci, http://dx.doi.org/10.3389/fnins.2013.00031
Jones, D. K., & Basser, P. J. (2004). Squashing peanuts and smashing pumpkins”: How noise distorts diffusion-weighted MR data. Magnetic Resonance in Medicine, 52(5), 979-933.
Diffusion is a pretty noisy technique anyway (low signal to noise ratio). From the acquisition standpoint you have noise coming from the pulse sequence used to collect it (typically a single-shot EPI), which can have spikes or distortions caused by sinuses and phase encoding direction. Also the b-value itself can introduce noise, as the higher the b-value the lower the SNR for both GM and WM structures. We try to compensate for some of these by collecting multiple averages of each direction... but the noise is substantial, and that's just on the collection side of things.
When it comes to actually estimating the tensors, the metric you are using (FA, ADC) can also have issues depending on a variety of factors impacting your data.
Some useful papers:
Soares, J. M., Marques, P., Alves, V., & Sousa, N. (2013). A hitchhiker's guide to diffusion tensor imaging. Front Neurosci, http://dx.doi.org/10.3389/fnins.2013.00031
Jones, D. K., & Basser, P. J. (2004). Squashing peanuts and smashing pumpkins”: How noise distorts diffusion-weighted MR data. Magnetic Resonance in Medicine, 52(5), 979-933.
Just to add to Dr. Molfese's excellent explanation, these sources of noise contribute to uncertainty or error in the vector directions of the diffusion tensor (or whichever model you might be using). Since these vector directions are the basis of determining streamlines during tractography, this error gets propagated and amplified as the streamline grows. Therefore, short streamlines have low error, while long streamlines are much more susceptible to the aggregation of these small sources of error. Probabilistic tractography methods quantify this uncertainty (to some degree), while deterministic methods do not.
Dear Nicholas,
Can you please explain how these errors can be quantified using probablistic methods and what's their importance in diffusion studies?
Thanks.
I should clarify a bit. Probabilistic approaches typically generate streamlines from a set of seed voxels and then estimate the probability that any other voxel X is connected to that seed based on how many streamlines pass through X. Therefore, noisier data will create more uncertainty in the streamlines, which will be represented by more "spread" in the probability estimates (i.e., truly connected voxels will have lower probabilities, spuriously connected voxels will have higher probabilities). Technically, these probabilities are also affected by the anisotropy of the regions through which they travel, so the uncertainty is not purely based on noise, but if one is comparing multiple acquisition schemes to determine which is the "noisiest", probabilistic tractography will provide this information. The importance to DTI studies is exactly that confound between anisotropy and noise. The following paper does a decent job of explaining the situation, though I know I've seen more recent discussions (I just can't remember where at the moment).
Article Characterization and Propagation of Uncertainty in Diffusion...
Thanks to you for the explanation..... I think we have more discussion. Such as, why there are still search by using DTI, though it has a problem in the areas of complex intra-voxel such as (crossing, kissing and fanning fibers). Notice that there are other techniques overcome these limitations such as HARDI, QBI and DSI.
Indeed, the term 'DTI' is now an outdated one. The currently preferred techniques are simply referred to as 'diffusion MRI', incorporating beyond the tensor model. e.g. as Mohammad pointed out, HARDI, QBI and DSI. If you have access to advanced techniques (i.e. you have a research key for your scanner), such as multishell and/or compressed sensing, then the SNR will be higher as will the angular resolution; moreso, applying techniques to the pulse sequence such as CAIPI/CAIPIRINHA will greatly reduce the g-factor. Acquiring a second dataset using your same pulse sequence but with reverse phase-encoding (reverse blipping) will also reduce artefact (you only need to acquire a single bzero and the minimum diffusion-encoded directions i.e. 6 for this second sequence as you will only be using the reverse-blipped bzero for the topup procedure). Processing the output of topup through eddy (and appropriately rotating the matrix i.e. bvecs) will greatly improve your analyses.
- Badges
- Science topic
- Similar topics
- Biological Science
- Neuroscience