This repository is dedicated to the group project of the course "Introduction to Deep Learning for Image Analysis and Computer Vision" at the Universeity of Luxembourg. 4 tasks that are asked in the main PDF of homework are solved and explained in this repository. The implementation and results of tasks 2 and 4 are also provided. In the current README, the solutions are documented and explained.
Please find the solution in the main PDF of homework, containing all the questions and solution of only task 1.
Please find the solution in task2.pdf. The code used to generate the plots of the PDF is provided in task2.m
In the following, a brief description of solution of Task 3 is provided: The initial images are displayed in the following:
The task3.m file contains the processing implementation of the Abingdon-Cross images. The first step of our processing is to denoise the noisy Abingdon-cross image; this is done using the pre-defined denoiser CNN architecture DnCNN, which can be used in the denoiseImage() function.
The architecture is represented in this picture:
In the following, the denoised version of images using the network are displayed:
As one can see, the denoising was rather successful.
After the denoising, the binary segementation is carried out using the imbinarize(), which successfully segements the image into "background" and "cross".
The resulting segmented images can be seen here:
The area and perimeters are computed using bwarea() and regionprops() repsectively, which returns a pixel count. The aforementioned quantities are presented in the following table:
| Area | Perimeters | |
|---|---|---|
| Binary Image | 3.5442e+04 | 1.2507e+03 |
| Denoised Image | 1.3488e+03 | 3.5347e+04 |
Finally, the DICE score is used to check how "far apart" both segmentations are. This score returns a value of 0.9947.
Cmake is a requirement for installing ANTs from source. Choose the Cmake version, and then use it in the wget command below
srun -p batch --time=24:00:0 -N 2 -c 12 --pty bash -i
wget https://github.com/Kitware/CMake/releases/download/v3.25.2/cmake-3.25.2.tar.gz
tar xzf cmake-3.25.2.tar.gz
cd cmake-3.25.2/
./configure --prefix=$HOME
gmake
gmake installInstructions of installing ANTs on HPC cluster, partly followed by this link:
cd medicimage/
git clone https://github.com/ANTsX/ANTs
mkdir antsbin
cd antsbin
ccmake ../ANTsNavigate into cmake and type c and then g then exit back to the
terminal.
make -j 4Navigate to the root folder contating ANTs and antsbin
cp ANTs/Scripts/* antsbin/ANTS-build/Examples/Now we need to set the bashrc paths accordingly:
nano ~/.bashrcAdd the following paths but adjust it to your own machine:
export ANTSPATH=/home/users/hvaheb/medicimg/antsbin/ANTS-build/Examples
export PATH=$PATH:$ANTSPATHDo the same steps with ~/.bash_profile.
Side Note: To copy output images from cluster to local machine, the following
scp -r iris-cluster:/home/users/hvaheb/medicimg/output/ /home/hamed/Documents/Projects/medical/dataset/output/The outpus of this section and next section are all stored in this Google Drive link.
Samples were chosen from the BRATS dataset to apply image registration using ANTs software.
A particular sample used is the BRATS_003.nii.gz file.
The header of nifti files derived from BRATS were checked to have more sense of their format. Using the value of sform_code or qform_code, the coordinate system of the file is determined.
The values for these codes are defined as following:
| qs form_code value: | (x,y,z) coordinate system refers to: |
|---|---|
| 0 | Arbitrary coordinates |
| 1 | Scanner-based anatomical coordinates |
| 2 | Coordinates aligned to another file's, or to anatomical "truth" |
| 3 | Coordinates aligned to Talairach-Tournoux Atlas; (0,0,0)=AC, etc. |
| 4 | MNI 152 normalized coordinates. |
The reg.py script is written to check for all nii files in a directory with name imagesTr and to append TRUE if the the qs-form had value 4, i.e., the image was registered with MNI template.
For all the images in BRATS dataset, the header's number was 4 i.e., they were scanner-based based coordinates.
In this sections, all nii files' visualizations are obtained from screenshots of the ITK-SNAP software.
The initial sample image BRATS_003.nii.gz is visualized in the following:
Evidenced by the figures, there is a bit of tilt in the horizontal direction, which is speculated to be attributed to the scanner-based (x,y,z) coordinate system, as investigated in Header. To restore image to a proper alignment space, the MNI template and segmentation template are used. The former is the standard template of MNI space, while the latter contains 3 big meta-brain structure, which is why it is called segmentation, as it contains the last segemented component of the former template.
Provided in the cluster is the MNI space coordinated image t1.nii, whic is used to transform the original image BRATS_003.nii to t1.nii, which is also called the moving image.
The antsRegistration commaned was used to map the original image to the MNI space, as follows:
antsRegistrationSyNQuick.sh -d 3 -f /scratch/users/ahusch/MSDS_19/MNI_SPACE/t1.nii -m /scratch/users/ahusch/MSDS_19/DATASETS/BRATS_dataset/imagesTr/BRATS_003.nii.gz -o /home/users/hvaheb/medicimg/output/003/BRATS_003_mapped -j 12
In below the outputs are presented:
Provided in the cluster is the MNI space coordinated image simple_segmentation.nii but it is segmented to have only the inner structure, which is used to transform the original image BRATS_003.nii to t1.nii, which is also called the moving image.
antsRegistrationSyNQuick.sh -d 3 -f /scratch/users/ahusch/MSDS_19/MNI_SPACE/simple_segmentation.nii -m /scratch/users/ahusch/MSDS_19/DATASETS/BRATS_dataset/imagesTr/BRATS_003.nii.gz -o /home/users/hvaheb/medicimg/output/seg/003/BRATS_003_mapped_seg -j 12
Using the nii filled registered (normalized) to MNI space in Image Restoration, this section is dedicated to segmenting tumor from other regions of brain.
The following syntax was attempted using Atropos, a segemntation tool from Ants:
Atropos -d 3 -a /home/hamed/Documents/Projects/medical/dataset/imagesTr/BRATS_003.nii -c 5 -i 'KMeans[5]' -o BRATS_003_seg
Atropos -d 3 -a /home/users/hvaheb/medicimg/output/seg/BRATS_003_mapped_segWarped.nii.gz -c 5 -m 'MLP-EM' -i 100 -k 'KMeans' -o /home/users/hvaheb/medicimg/output/segment/BRATS_003_segmented
There was not output, hence I proceeded with Python. Please find the Python notebook for segmentation.
In below the normalized nitfi image of BRATS_003 as well as its label (the tumor provided in the directory segmentation) are visualized.
Using Nilearn, the following segmentation plots are derived from the nii files of BRATS_003
Due to time limit, only one sample, which is BRATS_003 is considered for training set, and sample BRATS_200 is considered for test set.
The idea to quantify where tumors are more likely to be found could be handeled in the following manner:
Firstly, segment the images that are available into "Tumor" and "Non-Tumor" areas, using a binary segmentation.
After binary segmentation, one could transform them into grayscale images, with black being 0 and the former 1 (white) taking the value 255. This would be done for all images in the dataset. Now in order to get a general picture on where one would predominantly find the tumors, one could create an average of the voxels. This only works because we find oruselvess in the same, standardized MNI space.
This results to getting darker spots for an averaged image (over the whole dataset), in which one predominantly finds tumors, kind of like a grayscale version of a heatmap.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent