# Segmentation for Biomarker Extraction
###### tags: `CVC Lab` `Aaron - 2023`
summaries by Aaron Dominick
## Hack Overview
Covers the proposal for level-set segmentation for DaTSCAN/MRI. It will also cover similar techniques currently used within literature.
### 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)
9. Level-Set Segmentation

#### 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

# Previous Work (Chan-Vese Algorithm)
A level set is a set of points in the domain of a function where the function is constant. Starting with a domain $\Omega$, the curve or boundary of our desired segmented region can be described as zero level set of a smooth function $\phi$. This boundary is $\Gamma=\{\phi(x)=0\}$ and separates our domain into two regions, $\Omega_{in} = \{\phi(x)>0\}$ and $\Omega_{out} = \{\phi(x)<0\}$ where x is a point in the domain $\Omega$. Letting I(x) being an image defined on the domain, the optimal separation is the one that follows the following optimization problem: $$\min \text{E}(\phi,c_1,c_2) = |\Gamma| + \int_{\Omega_{in}} |I(x)-c_1|^2dx + \int_{\Omega_{out}} |I(x)-c_2|^2dx$$ Where $|\Gamma|$ is the length of the segmenting curve, $c_1$ and $c_2$ are the average intensities inside and outside the segmenting countour: $$c_1 = \frac{1}{|\Omega_{in}|}\int_{\Omega_{in}} I(x)dx, \space \space c_2 = \frac{1}{|\Omega_{out}|}\int_{\Omega_{out}} I(x)dx$$
**Figure** Demonstration of level set segmentation. Changing the value of the zero set can transform the region.

The evolution of the level set is governed by the following PDE: $$\frac{\partial \phi}{\partial t} + V|\nabla \phi| = 0$$ where V is the velocity field for image segmentation

# Segmentation using other techniques.
## Data and Data editing
I found a training/testing set of 101 T1 weighted MRI + brain segmentations of ~50 regions assembled by mindboggle.(https://mindboggle.info/data.html). The goal of this section was to train a model to take in T1 weighted MRI and output a segmented image for the Caudate/Putamen regions. Both the training data and the example PPMI data undergo the same preprocessing done by a program known as Freesurfer. Freesurfer is a non-machine learing segmentation program. One main advantage of Freesurfer is the use of a probabilistic atlas that uses location to boost classification. (https://surfer.nmr.mgh.harvard.edu/ftp/articles/fischl02-labeling.pdf) The following steps are done for freesurfer segmentation and some of the steps are done for MRI preprocessing for our models.

## PPMI Data Preparation
There is no universally accepted procedure for the preprocessing and standardization of MRI images for machine learning. Some work has been done to determine the most optimal set of steps [https://www.nature.com/articles/s41598-020-69298-z]. Following that advise there are two steps which will be applied to our images before use in the Segmentation algorithm. The first step is Z-score normalization for standardization of images across different machines. The second step is absolute discritization for clear grouping of pixels values. This will reduce the effect of noise on an image as well. Random noise will be applied to the training set to build a more robust model.

## Segformer
Segformer is a semantic segmentation framework which unifies transformers with lightweight multilayer perceptron (MLP) decoders.
**Figure** The Segformer architecture is comprised of an encoder and decoder.

### SEGFORMER RESULTS ON MINDBOGGLE
**Figure** Segformer after on testing set.
## UNETR
UNETR consists of a transformer encoder that directly utilizes 3D patches and is connected to a CNN-based decoder via skip connection.
https://github.com/Project-MONAI/research-contributions/tree/main/UNETR/BTCV
**Figure** UNETR model

### UNETR RESULTS ON MINDBOGGLE
**Figure** UNETR after on training set.


**Figure** UNETR after on validation set.


### RESULTS
Segformer preforms much better than UNETR when it comes to the segmentation of specifically the putamen and caudate regions of the brain.

### RESULTS on PPMI Data
**Figure** Results of Segformer from patient 4005

The results from Segfomrer on data from PPMI has been promising. Some errors still occur, however a rough map of the Caudate and Putamen region can be constructed in ~5 minutes compared to the hours it takes on a non-machine learning program such as Freesurfer.

### Analysis of Volumes

### Final Results

### References
[1]https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=902291
[2]https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5885990/
[3]https://cs.uwaterloo.ca/research/tr/2008/CS-2008-12.pdf
[4]file:///oden/adominick/Downloads/IC.3.ICSIVP2012-MedicalImageSegmentationusingLevelsetMethodwithoutreinitialization.pdf
[5]https://arxiv.org/pdf/1107.2782.pdf