# DaTscan Imaging Overview
###### tags: `CVC Lab` `Aaron - 2023`
`
summaries by Aaron Dominick
## Hack Overview
This hack is a collection of imporant information related to DaTscan imaging for Parkinson's Disease. It has 5 key sections:
1. DaTscan Overview
2. DaTscan Biological Importance
3. DaTscan Reconstruction
4. DaTscan Previous Work
5. Datscan PPMI Data
### Relevant Hacks
Aaron Dominick PHD Timeline: https://hackmd.io/Xawu3zWwS8-Zw5f6ivsfMg
#### Longitudinal Data:
1. Current Longitudinal Methods for Parkinsons's (https://hackmd.io/HFy0yiXlQWKVeRueCrrckQ)
2. Trajectory Net (https://hackmd.io/-6FLIziJQ7qeAIpsXj193A)
3. Longitudinal VAE (https://hackmd.io/GlmwNqwBQoa0aE8HA8N4Dw)
4. ODE2VAE (https://hackmd.io/e5phRKlqT2CW_13FCqWW2Q)
5. Gaussian Process Prior VAE (https://hackmd.io/JuFhjzz0TaG60l-i9xBecw)
6. iLQR-VAE (https://hackmd.io/QySMkFamSgqvH8UgU7QBHQ)
7. Evaluating MRI Data on Longitudinal Methods (https://hackmd.io/GY9ONkAGTe2icDwJe-yHrw?both)
8. Evaluating DaTscan Data on Longitudinal Methods (https://hackmd.io/jc3fHQWvSh2A8KOvK8lKRQ)

#### Parkinson's Data and Previous Methods:
7. Actionable Intelligence: Understanding and Treating Parkinson’s Disease (PD): https://hackmd.io/@GeoZhai/B1yB7iz45
8. PPMI dataset exploration: (https://hackmd.io/BBvuXZ8ZTn6VqcmFYCk4_w?both)
9. Patient X's Data:(https://hackmd.io/rTBOG_rTRAquOXqGC_JWaw)
10. DaTscan Data: (https://hackmd.io/rs2D-YBjR3WG_xCnCjC6EA)
11. Parkinson's Biomarkers Overview (https://hackmd.io/yx6iwA0wSu2ephw07Q56SA)
12. Parkinson's Insights and Existing ML Methods for Diagnosis
https://hackmd.io/G3UvWxbuSXCHkZ0c9p_lew?both

## Questions
1. ~~State of the Art for Extraction(input to biomarkers)~~
2. ~~How is the functional and anatomical synchronized~~
3. ~~What method is used to measure the QOI~~
5. ~~Correlation between biomarker and disease~~
6. ~~State of the art for building, alligning, normalizing DaTscan~~

Take DaTscan and learn biomarkers
## Two Goals
### Goal 1: Extraction of Region Specific and Patient Specific SBR(Or similiar metric)
**Input:** 3D DaTSCAN Images(109x91x91), where each pixel is 3.9 mm spaced and holds a value derived from the specific binding ratio.
**Optional Input:** 3D T1-weighted MRI for possible segmentation of different regions
**Outputs:** SBR from the caudates, putamen(total, posterior, and anterior), and total striatums. Possible shape metrics?

The putamen was subdivided into anterior and posterior parts using a marker placed on the coronal plane at the level of the anterior commissure.

#### What separates the regions in MRI?
The MRI images provided are of varying dimensions. The example below is (176x240x256) with pixel values within the range 0-256 representing black and white.
### Goal 2: Longitudinal Analysis of DaTSCAN Markers

## Key Biomarkers for DaTCAN
> The result of DaTscan imaging was defined “abnormal” when a reduced SBR (13), an abnormally increased AI (13), or abnormal patterns such as an egg shape (bilateral reduction of tracer uptake in the putamen and normal or almost normal uptake in caudate nuclei) (grade 2) and a mixed type (normal or almost normal tracer uptake in bilateral caudate nuclei with asymmetrical tracer uptake reduction in the putamen of one side) (grade 3) (12) were recognized by visual assessment. [8]
1. An asymetric decrease in tracer uptake between the left and right striatum
* quantified by asymmetry index(AI) where RS and LS is right striatal SBR and left striatal SBR: with the normal range: <11.05%. [7] $$\textbf{AI} = \frac{RS-LS}{RS+LS}*200$$

2. Total decrease in binding (Reduced SBR) in one or both of the putamen
* In literature this is typically done by assigning an intensity value(in SBR - calculated below) to the left caudate, right caudate, left putament, and right putamen. Putament is the first to disapear resulting in the shape change
* **The region that has shown the most useful for diagnosis is the posterior putamen [14].** The SBR of this region was able to distinguish between normal and abnormal scans with a ~90% sensitivity and specificity depending on the data set used. Other areas such as striatum, total putamen, and anterior putamen had similar results for certain datasets but decreased on others making posterior putamen the most reliable. They recommended threshold values for DaTQUANT posterior putamen are: SBR of ≤1.0, z-score of ≤−1.8, and percent deviation of ≤−0.34. DaTQUANT is a common software used for analysis and is talked about later.

**Things to note**
* Correlation between DaTscans and MDS-UPDRS(motor function) scores only occurs when split into progression subgroups otherwise there is no correlation
* Decline in DAT is ~6.3% per decade on average for both sexes, with a range of 3.6%-7.5%, the rate has been shown to be linear with both decrease in DAT in the striatum and neuronal loss.
**Figure** Three PD patients visualzed. Patient A has circular patter on both striatum and an asymmetric reduction. Patient B has circular shaped on one end and C has normal shape with asymmetric decrease in one side (caudate + putamen). All three are in the early stages of PD classified as an MDS-UPDRS < 4

## Datscan Overview
DaTscan imaging is one of the most common types of imaging done for Parkinson's disease (PD) and is typically used to confirm a doctor's diagnosis. A radioactive agent(Ioflupane I 123) is injected into the patient’s bloodstream and is tracked using SPECT — a noninvasive scanner that uses radiation detectors to image the cells that take up the agent and emit radioactivity. DaTscan specifically marks the brain cells that carry dopamine transporter(DAT). The higher the number of healthy cells having the transporter protein present, the brighter the images are. In this way, the intensity of the brain images determines the health of the brain cells. More specifically. a hallmark for PD in DaTscan is an asymmetric reduction in tracer uptake in the caudal to rostral gradient within the striatum(caudate and putamen) where signal loss is the most prevalent. This asymmetry is shown in the figure below:

**Figure** Reconstructed DaTscan image of a normal brain and someone with Parkinsons. Note the reduced signal and asymmetry in the PD patient.

However, the images we have are still in their raw form. During imaging, the radioactive agent releases gamma rays which are picked up by a gamma detector. 2D images of 3D areas are captured at multiple angles allowing for reconstruction of a 3D image. This image is given in the form of a sinogram for each respective slice. PPMI gives the DaTscans in the raw sinogram form. A sinogram is comprised of two dimensions. The vertical dimesion is the rotation angles of the scanner. The horizontal axis is the offset of the scanner beam. If a gamma ray is detected at that rotation/offset pairing, then an event is counted. The values stored within the sinogram matrix represent the number of events detected (integers).


**Figure** DaTscan image being stored into sinogram form[1]
## DaTscan Biological Importance
The striatum and substantia nigra are tied in function. Both are located in the basal ganglia within the midbrain and play an important role in both the brain's reward and movement systems, including motor function, action planning, descision-making, motivation, reinforcement, and reward perception. The basal ganglia is connected to the motor cortex and the thalamus through inhibitory and excitatory synapses. When a person wishes to complete a voluntary movement, the motor cortex sends an excitatory signal (RED ARROW) with the release of glutamates into the stiatum which then inhibits (BLUE ARROW) the globus pallidus internal(GPi) which in turn excites the thalamus. The thalamus then sends an excitatory signal back to the motor cortex exciting the muscle. The substantia nigra(SNpc) plays a regulatory role in this process. It sends dopamine through the niagrostiatal pathway(PINK ARROW) to D1 dopamine receptors in the striatum. The increase in dopamine increases the inhibitory effect on the GPi.
* **Direct Pathway** Cortex $\rightarrow$ Striatum $\rightarrow$ GPi $\rightarrow$ Thalamus $\rightarrow$ Cortex
* Controls active muscle movement

There also exists an indirect pathway in the basal ganglia that prevents undesirable movement. The excitation of the striatum also inhibits the globus pallidus external(GPe), which then inhibits the subthalamic nucleus(STN). The STN then excites the GPi, inhibiting the activity of the thalamus. The substantia nigra also plays a role in this indirect pathway. It sends dopamine to bind to D2 receptors in the stiatum, lowering the inhibitory ability of the striatum.
* **Indirect Pathway** Cortex $\rightarrow$ Striatum $\rightarrow$ GPe $\rightarrow$ STN $\rightarrow$ GPi $\rightarrow$ Thalamus $\rightarrow$ Cortex
* Prevents undesirable movement

Parkinson’s disease is caused by a substantial loss of dopaminergic neurons in the substantia nigra(specifically the pars compacta), effecting the nigrostiatal pathway and leading to the depletion of dopamine in the striatum(caudate,putamen). This eliminates the regulatory effect that the substantia nigra plays on the pathway which results in the motor symptoms that define PD. Parkinsonian motor symptoms do not appear, however, until up to 50–80% of pars compacta dopaminergic neurons have died. [5]


**Figure** (a) An MRI of the brain. The striatum is marked in red featuring 3 key areas. The caudate (top), putamen(lower right), and globus pallidus (lower left)[2] (b) The interactions present within the basal ganglia. The striatum connects to the global pallidus(GP exterior/interior), then to the subthalamic nucleus(STN)

## DaTscan Reconstruction(WIP)
A number of reconstruction and filtering techniques are applied to this image before information can be visually extracted. The brief process is laid out in PPMI's SBR calculation below.

**Preprocessing of Images**
> SPECT raw projection data was imported to a HERMES (Hermes Medical Solutions,
Skeppsbron 44, 111 30 Stockholm, Sweden) system for iterative (HOSEM) reconstruction. This was done for all imaging centers to ensure consistency of the reconstructions. Iterative
reconstruction was done without any filtering applied. The HOSEM reconstructed files were then transferred to the PMOD (PMOD Technologies, Zurich, Switzerland) for subsequent processing. Attenuation correction ellipses where drawn on the images and a Chang 0 attenuation correction was applied images utilizing a site specific mu that was empirically derived from phantom data acquired during site initiation for the trial. Once attenuation correction was completed a standard Gaussian 3D 6.0 mm filter was applied. These files were then normalized to standard Montreal Neurologic Institute (MNI)space so that all scans were in the same anatomical alignment.

**Current Methods for Extracting SBR Done by PPMI**
> Next the transaxial slice with the highest striatal uptake was identified and the 8 hottest striatal slices around it were averaged in to generate a single slice image. Regions of interest (ROI) were then place on the left and right caudate, the left and right putamen, and the occipital cortex (reference tissue). Count densities for each region were extracted and used to calculate striatal binding ratios (SBRs) for each of the 4 striatal regions. SBR is calculated as (target region/reference region)-1.

Starting off, ordered subsets expectation maximization. In this algorithm, the set of all projections {1...I} is divided into a series of subsets $S_{t}$,t = 1...T. Usually, these subsets are exhaustive and non-overlapping, i.e. every projection element i belongs to exactly one subset $S_{t}$. The data y is usually organized as a set of parallel projections, indexed by the projection angle $\phi$. The subsets are then produced by assigning all the data for each projection angle into individual subsets. The algorithm goes as follows:

**Figure** Algorithm for calculating OSEM

Where A is the system matrix(a matrix consisting of the probability that a unit of radioactivity in j gives rise to the detection of a photon in line of response i) i are the sinogram elements and total the number of rotation angles (vertical dimension in above sinogram). j is the number of image voxels or offset within scanner(horizontal dimension in the above sinogram). b is the additive reconstruction term.


## DaTscan Previous Work
A metric known as the specific binding ratio (SBR) is typically assigned to each voxel through the following equation: $$SBR_{voxel} = (Intensity_{voxel} - \mu)/\mu$$

Where $\mu$ is the mean or median intensity in a reference region, (such as the occipital lobe) that doesn't have specific ligand binding.

SBR is not the only value used for evaluation of DaTSCANs, binding potential index and total binding potential index are names of similiar parameters that evaluate the difference between the intensity of the striatum and intensity in a non-active region.

### Current use clinical DaTSCAN ROI/VOI Extraction [11]
The regions of interest include the caudates, putamens, and whole striatums. There are two main methods for metric extraction. The ROI extraction makes a 2D representation from the 3D image for simplified extraction. VOI extraction takes advantage of full 3D DaTSCAN image and structural images (MRI). **A program known as DaTQUANT is commonly used to preform**
1. **Manual ROI extraction** In current literature, including PPMI, the ROI is manually deliniated by an expert
2. **Template ROI extraction** This section encompasses two methods where templates are utilized
1. The first involves placing preconstructed-trapezoidal ROIs over the striatal compartment of the 2D image [13]
1. The second involves taking the DaTSCAN image and nomalizing it to a space where pre constructed ROI are known and can be placed over the modified image [12].
3. **VOI extraction** This method involves the use of complimentary structural imaging to determine the volume of interest in the DaTSCAN. The VOIs are segmented out of structural images such as MRI manually through segmentation software (ITK-snap). Then the DaTSCAN and MRI are coregistered using **rigid body transformation** into the same space.
**Figure** The three primary techniques for extracting the ROI/VOI from DaTSCAN images

### Extraction of Parameters through State of the Art Automated Means
#### A Novel Automatic Approach for Calculation of the Specific Binding Ratio in [I-123]FP-CIT SPECT [9,10]
Extracted SBR in an automated method using Trapezoidal VOI. Starting off, the process identified the maximum voxel in an average image of all the slices containing the stiatum. After which the point was mirrored across the image mid section and used to construct an trapezoidal volume of interest using the distance between the maximum volume points. **This just highlights the region where the striatum is present**

The volume of the striatum is then extracted by the known average striatum volume of (11.2 mL). Using the volume of the voxel (2.2mm^3), they calculated the number of voxels included in the striatum (1052 voxels). The area was then extrapolated to all striatum slices and the top 1052 consecutive voxels for both left and right striatum were extracted out to match the average striatum volume.
**Figure** Automated extraction of SBR values using trapezoidal volume of interest to create a mask of the striatum

#### CADA—computer-aided DaTSCAN analysis [6]
They employed a gaussian mixture model segmentation routine to extract two binary masks corresponding to the left and right striata. An expectation maximization model estimates the mixed parameters. Subsequently, they fitted an ellipsoid surface on each mask using a non-linear optimization algorithm and then used the ellipsoid space to extract a set of features characterizing the basal ganglia. From this, they generated quantitative measures based on striatal intensity, shape, symmetry and extent.
**Figure** CADA workflow

#### DaTQUANT [15]
DaTQUANT is a software used to analyze DaTSCAN images by SBR reading in different regions of interest. The process goes as follows:
1. First, each brain volume was spatially normalized to a standard geometry.
2. **Locating the Central Slice:** The central plane through the subject’s striatum is found by locating the maximum value in the top-to-bottom profile in the summed sagittal view of the reconstructed images.
3. **Summing the Central Striatal Slices** A parameter N describes the number of slices that are summed to place the 2-dimensional regions of interest (ROIs). In order to maintain symmetry, the same number of slices is added above and below the central slice through the striatum.
4. **Locating the Individual Striata** Within the 2-dimensional transverse summed slice, by looking for maxima in the subject’s left-to-right profiles, the intensity centroids of the left and right caudate are located.
5. **ROI Placement** A crescent-shaped ROI is automatically placed on the occipital area of the brain. A caudate ROI (green circle) is moved automatically within the vicinity of each located striatum until the SBRquant locates a maximum count value representing the initial location of each caudate. The smaller caudate ROI (white circle) is then placed by the software in its center and is used for the caudate’s count density measurement.
6. **Locating the Left and Right Putamen** The putamen ROI (white circle) is automatically determined in relation to the caudate (white circle) using the parameter d in units of pixels. The SBRquant software allows the selection of different values for d from its default settings, thereby allowing the angular placement of the putamen ROI to be varied around the caudate systematically for all subjects in the group being analyzed until a maximum value is found for extracting the putamen’s count density.
7. SBR is then calculated using the above formulat

**Figure** How the SBR values are extracted from DaTSCAN image using a semi-quantitative method.

### How DaTSCAN Relates to Patient State
DaTscan shape progression may provide useful insight into the progression of the disease. Specfically, the striatum's initial comma shape transforms into a circle or disappears due to dopamine levels decreasing in the putamen first. This can be seen in the image below.

Work has already been done to extract additonal disease progression markers from DaTscan images [2]. Some interesting findings from the paper is the presence of 3 different progression subtypes whose progression is correlated with patient state measured by UPDRS scores. They identified imaging patterns present in these subtypes that represent progression speeds. They showed that the loss of asymmetry in the disease is the fastest spatial progression pattern while mean SBR being the slowest. This paper only took into account the DaTscan images and not any clinical data as covariates.

Another paper applied machine learning to further classify these subtypes using clinical data and DaTscan. They were focused on separation into three separate subclasses, but the covariates they used may still be useful for us to consider.

**Figure** DaTscan image of a normal brain and someone with Parkinsons. Note the reduced signal and asymmetry in the PD patient. Note also the shape change from a comma shaped signal to a circle.

**What is shown is both in DaTSCAN images and Parkinson's progression there are ~ three unique substypes that progress at different speeds** [13]
## DaTscan PPMI Data
The resolution of the images has a range of (128,128, 120/240/480) with a pixel spacing in the range of 2.6-3.3 mm$^{2}$. **DaTscan images are from 2 cohorts, control and PD**. For PD there are a total of 2676 scans. For control there was 276 scans.
* If we dont account for overlap of days there are: 368 patients 1 scan, 112 patients 2 scans, 125 patients 3 scans, 219 patients 4 scans, 34 patients 5 scans, 42 patients 6 scans, 11 patients 7 scans, 1 patient 8 scans, 3 patients 9 scans, 3 patients 10 scans, 2 patients 11 scans, 0 patients 13 scans, 3 patients 14 scans, 3 patients 15 scans. Howevere a lot of these higher scan numbers are overlap at the same time.
* If we account for time unique scans, then there are 381 patients with 1 scan, 125 patients with 2 scans, 154 patients with 3 scans, 273 patients with 4 scans, 9 patients with 5 scans, and 2 patients with 6 scans.
* There were 265 non overlap control scans all with a unique patient.

We also have access to analysis done upon the DaTscan images by PPMI, specifically calculation of SBR in the left/right caudate andthe left/right putamen.

> SPECT raw projection data was imported to a HERMES (Hermes Medical Solutions,
Skeppsbron 44, 111 30 Stockholm, Sweden) system for iterative (HOSEM) reconstruction. This was done for all imaging centers to ensure consistency of the reconstructions. Iterative
reconstruction was done without any filtering applied. The HOSEM reconstructed files were then transferred to the PMOD (PMOD Technologies, Zurich, Switzerland) for subsequent processing. Attenuation correction ellipses where drawn on the images and a Chang 0 attenuation correction was applied images utilizing a site specific mu that was empirically derived from phantom data acquired during site initiation for the trial. Once attenuation correction was completed a standard Gaussian 3D 6.0 mm filter was applied. These files were then normalized to standard Montreal Neurologic Institute (MNI)space so that all scans were in the same anatomical alignment. Next the transaxial slice with the highest striatal uptake was identified and the 8 hottest striatal slices around it were averaged in to generate a single slice image. Regions of interest (ROI) were then place on the of left and right caudate, the left and right putamen, and the occipital cortex (reference tissue). Count densities for each region were extracted and used to calculate striatal binding ratios (SBRs) for each of the 4 striatal regions. SBR is calculated as (target region/reference region)-1.

In addition, there was a visual interpretation done by experts in the field on whether a DaTscan had normal or abnormal signs.

> Abnormal images typically fall into at least one of the following three general categories:
a) Activity is asymmetric, e.g. uptake in the region of the putamen of one
hemisphere is absent or greatly reduced with respect to the other. Uptake is still
visible in the caudate nuclei of both hemispheres resulting in a comma or
crescent shape in one and a circular or oval focus in the other. There may be
reduced uptake between at least one striatum and surrounding tissues.
b) Ioflupane uptake is absent in the putamen of both hemispheres and confined to the
caudate nuclei. The signal is relatively symmetric and forms two roughly circular or
oval foci. Uptake of one or both is generally reduced.
c) Uptake is absent in the putamen of both hemispheres and greatly reduced in one
or both caudate nuclei. Uptake of the striata with respect to the background is
reduced (1).

## MRI Segmentation of Striatum
**Figure** MRI of PPMI patient
If we utilize MRI to help us in region separation and identification for DaTSCAN, there have been a few techniques developed.

In the the case of machine learing. Two convolutional neural networks were used to segment the areas into striatal regions (1) or non-striatal regions (0). [S1]
Free Surfer is a commonly used-open source software used for automated segmention in neuroscience domains. [S2]
## References
[1] https://humanhealth.iaea.org/HHW/MedicalPhysics/NuclearMedicine/ImageAnalysis/3Dimagereconstruction/index.html
[2] https://en.wikipedia.org/wiki/Striatum
[3] https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=9225726
[4] https://www.nature.com/articles/s41598-018-37545-z
[5] https://openscholarship.wustl.edu/cgi/viewcontent.cgi?article=1585&context=eng_etds
[6] https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4754234/
[7] https://pubmed.ncbi.nlm.nih.gov/29478082/
[8] https://www.frontiersin.org/articles/10.3389/fneur.2021.656679/full
[9] https://pubmed.ncbi.nlm.nih.gov/32397547/
[10] https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6314989/
[11] https://ejnmmiphys.springeropen.com/articles/10.1186/s40658-022-00519-2
[12] https://pubmed.ncbi.nlm.nih.gov/22072699/
[13] https://www.nature.com/articles/s41531-022-00439-z
[14] https://www.mdpi.com/2379-139X/7/4/81#B14-tomography-07-00081
[15] https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7257254/

https://pubmed.ncbi.nlm.nih.gov/33892414/

ML:
https://www.degruyter.com/document/doi/10.1515/jisys-2018-0261/html?lang=en

[S1] Segmentation Database mention in a paper
https://www.sciencedirect.com/science/article/pii/S0165027016302321?via%3Dihub

[S2] https://surfer.nmr.mgh.harvard.edu/ftp/articles/fischl02-labeling.pdf