Differential tractography for longitudinal data


Figure: conventional tractography compared with differential tractography. Conventional tractography tracks "existence", whereas differential tractography tracks "differences".

Tracking differences instead of tracking the existence
Differential tractography is a new type of tractography that compares repeat scans of the same individuals to capture neuronal injury reflected by a decrease of anisotropy. It is realized by adding a criterion to track along trajectories only if a decrease of anisotropy is found between repeat scans. Integrating this “tracking-the-difference” paradigm into the fiber tracking process results in a new tractography modality called differential tractography. It will track the exact portion of pathways exhibiting substantial differences in anisotropy. The additional criterion ignores unaffected regions and enhances meaningful findings related to neuronal injury. 

Conventional fiber tracking, in comparison, is based on a “tracking-the-existence” paradigm. It only considers anisotropy from one MRI scan and thus will include all existing pathways regardless of whether they have an injury. 

The following are steps to reproduce the result published in the differential tractography paper. 

*This feature is only available after the 8/23/2019 version. Please update DSI Studio before you start.

Step 1: Experiment design

You will need to have "longitudinal data" (one baseline scan and one follow-up scan) of the same subject to run differential tractography. The scanning interval has to be more than one month. The optimal is more than 3 months because structural change needs time to build up. For estimating FDR, the optimal experiment setting include an additional "sham" scan, which is an additional baseline scan one or two days after the baseline scan. This additional baseline scan will be used in estimating the false discovery rate. If there is no sham scan, then an "alternative sham" approach can be used without the need for additional scans.

Step 2: Diffusion MRI acquisition

Differential tractography can be applied to any diffusion data set, including DTI data, multi-shell data, and DSI data. But, we found that higher b-value signals will be more sensitive to early-stage neuronal changes, whereas low b-value signals may include a lot of physiological fluctuations. If your data are acquired at a low b-value (e.g., < 3000), then you can expect to have a higher false discovery rate (FDR). In the original study, we used 256-direction grid sampling with b-max=4000 or even up to 7000 to get excellent FDR values lower than 0.05. Using DTI data may increase FDR to 0.2

The recommend b-table and acquisition setting can be found here. The setting in the link has b-max=4000. For differential tractography, you may need to increase both the voxel size to 2.5 mm cubic and b-value to 7000.

Step 3: Processing DICOM/NIFTI

After you collected the data, first convert DICOM/NIFTI to SRC fileThe webpage also has a quality check section. Make sure that all your SRC data are created and pass the quality control.

Step 4: Compare differences 

Open the baseline SRC file in [Step T2 Reconstruction], make sure that the mask at Step T2(a) is okay.

At [Step T2b(1)] Select [GQI] and set length parameter 1.25 (human)  or 0.6 (animal). The length parameter can be increased or decreased depending on the data quality. A higher value is more sensitive to crossing fiber but also more sensitive to noise.

At [Step T2b(2)] Click on the [Compare SRC...] button to assign the SRC file of the follow-up scan. 

Then click [Run reconstruction], and DSI Studio will generate a FIB for further processing

*TIP for batch processing: If you have a large amount of SRC files to process, you can place all SRC files together in the same directory and select all baseline SRC files when clicking on [Step T2 Reconstruction]. DSI Studio will load the first SRC file, and please select its matching SRC file. For the rest SRC files, DSI Studio will then do the same by findings the best matching follow-up scans.

Step 5: Quality control

(1) Check R2 value:

DSI Studio will generate a fib.gz file with a file name such as baseline.src.gz.odf8.f5.df.follow_up.R96.rdi.gqi.1.25.fib.gz
Here R96 means the R squared value is 0.96, which is very good. If the value is lower than R80, then you will need to check the data quality or email Frank to see if there is any problem with the data registration.

(2) Check baseline and follow-up FA map

a. Open the generated FIB file in [Step T3 Fiber Tracking]
b. At the top of the 3D window, switch the slice between "base_fa" (FA map of the baseline scan) and "study_fa" (FA map of the follow-up scan) to see if the alignment is good. If you see a large misalignment, please email Frank.

Step 6: Initial tractography inspection

Before invoking differential tractography tracking the differences, we will need to make sure that regular fiber tracking produces good quality of whole-brain tracks using following the steps:

a. Open the generated FIB file in [Step T3 Fiber Tracking]
b. In the right upper corner [Step T3c Option] window, expand the [Tracking Parameters]
c. At [Tracking Parameters][Terminate If], assign 100,000 "seed"
d. Click on the [Step T3d Tracts][Fiber Tracking] button to generate whole-brain tractography. Good quality of whole-brain tracks should look like this
e. Examples of poor-quality whole-brain tractography:

[Step T3c Options][Tracking Parameters][Tracking Threshold] is too low (left) or too high (right). 
Solution: adjust the threshold

Uncorrected EPI distortion
Solution: (1) correct distortion using opposite phase encoding data, or (2) assign a seed region in [Step T3a] that excludes anterior frontal lobe

Step 7: Tracking differences 

Using the parameters from Step 5, we can now track the differences:

a. In the right upper corner [Step T3c Option] window, expand the [Tracking Parameters]
b. At [Tracking Parameters][Terminate If], assign 1000,000 "seeds." This value can be increased to get more results.
c. At [Tracking Parameters][Differential Tracking Index], assign "dec_qa" to track decrease of "qa". You may use "dec_fa" to track the decrease of "fa" for DTI data. The increased anisotropy can be tracked by "inc_qa" or "inc_fa". 
d. At [Tracking Parameters][Differential Tracking Threshold], assign 0.1 to track anisotropy decrease larger than 10%. 
   The recommended value for "fa" is 0.1~0.3 and "qa" is 0.2~0.5. The value depends on the pathological condition. Lower values are for early demyelination, whereas higher values are for the axonal loss.
e. (Optional but recommended) You can use [Edit][Pruning (TIP)] to eliminate noisy results. I will repeat TIP for 5~10 iterations.

f. (Optional) The sensitivity and specificity can be adjusted by [Tracking Parameters][Min Length (mm)]. Lower value like 20 mm is more sensitive, whereas 40 mm is more specific. See the figure below for an example.
g. (Optional) You can add ROI from an atlas at Step T3a to limit the findings to a specific region.

Differential tracking will use the fiber orientations from the first scan to guide the fiber tracking algorithm, and the second scan is only used to provide the differences in FA.
Therefore if the baseline and follow-up studies are swapped, the results won't be the same.

*To recognize the anatomical structure of the findings, you can use findings as ROIs by [Tracts][Tracts to ROI] to run fiber tracking and then use [Tracts][Miscellaneous][Recognize Track] to determine the anatomical structures.

Step 8: Estimating FDR 

There are two ways to design experiments for sham scans. If there is no scam scan, the "alternative sham" method can be used to calculate FDR.

Approach 1 (Recommended)

Acquire baseline scans for patients and follow-up scans at 3 months (or longer). The sham scans can be acquired on the same patients on the same days as the baseline scans or a few days later. Do not scan sham scans right after baseline scans because this may falsely reduce FDR.

FDR= (total number of differential tracks count from sham scans) / (total number of differential tracks count from the follow-up scans)

Approach 2
Acquire baseline scans and follow-up scans for patients. Recruit healthy control subjects and do the same. The age/sex and all other settings should be matched between the patient and the healthy control group. 

FDR= (averaged number of differential tracks count from healthy controls) / (averaged number of differential tracks count from the patient groups)

The FDR calculation procedure using a sham scan is shown by the figure above:

The idea here is to compare the follow-up scan with a sham scan using the identical setting (*NOTE* make sure that the seeds count, minimum track length, and the number of TIP iterations are identical). 

The sham scan should generate fewer results than the follow-up scans. The FDR is calculated as (the number of tracks in the sham scan) divided by (number tracks in the follow-up scans)If the FDR value is lower than 0.2, then it is worth reporting. FDR < 0.05 is confirmative of the findings. If the FDR is not ideal, you can change the [Tracking Parameters][Differential Tracking Threshold], [Tracking Parameters][Min Length (mm)], or use more [Edit][Pruning (TIP)] iterations. FDR already considers multiple comparisons, and there is no risk of inflating the significance.

Alternative sham:

If there is no sham scan, we can use an alternative sham approach, but it will increase FDR and can make results not significant. 

The idea of an alternative sham is to use "inc_fa" as a sham for "dec_fa" (or vice versa). For example, we can assume that a patient should not have increased FA due to the disease, and thus any finding with increased anisotropy can be viewed as false-positive results.

you can change [Differential Tracking Index] to "inc_qa" (or "inc_fa" for DTI data) in the previous step as the "alternative sham".

For results from "dec_fa", the FDR is then calculated as (number of tracks in "inc_fa") divided by (number tracks in "dec_fa").
For results from "dec_qa", the FDR is then calculated as (number of tracks in "inc_qa") divided by (number tracks in "dec_qa").

Using other metrics

Differential tractography can be generated using other metrics. You will first need to prepare two nii.gz files of a metric that will be compared (e.g. DKI1.nii.gz and DKI2.nii.gz)

Use [Analysis][Add Tracking Metrics] and input DKI2-DKI1 to track the increase of DKI from the DKI1 scan to DKI2 scan, (for tracking the decrease of DKI, use DKI1-DKI2)

A new metric will be added to [Step T3c: Options][Tracking Parameters][Differential Tracking][Metrics] that allows you to track results.

Batch Processing Using GUI

DSI Studio supports GUI-based batch processing. For differential tractography, place all SRC.gz files in one folder, including the baseline and follow-up scans. Then name the SRC files in a way that the same subject's data share most of the prefix of the file name (e.g. sub001a.src.gz for baseline and sub001b.src.gz for follow-up). Then select all baselines SRC files in [Step T2 Reconstruction] and follow the same procedure mentioned above. DSI Studio will apply the same processing to all other data (if not, please report the issue in the discussion forum).

Video Tutorials

Quality control for differential tractography

Differential tractography on TBI patient

Differential tractography comparing healthy control and patients

Differential tractography on a stroke patient

Differential tractography combined with DKI on a stroke patient


1. Can I apply differential tractography to DTI dataset?
Ans: Yes, the exact procedure can be applied to DTI data, but the FDR will not be as good as the recommended acquisition (e.g., Diffusion MRI acqusition)

2. Can I apply differential tractography to study increased connectivity? 
Ans: Yes, but you may need a control subject or a sham scan to estimate FDR. You cannot use the "alternative sham" approach (see the original paper for details).

3. Can I combine differential tractography with ROI?
Ans: Yes, differential tractography can be combined with an ROI to limit the findings to a region.


Yeh FC, Zaydan IM, Suski VR, Lacomis D, Richardson RM, Maroon J, Barrios-Martinez J. Differential Tractography as a Track-Based Biomarker for Neuronal Injury. Neuroimage. 2019.