This repository currently consists of only the pre-processed ADNI data along with the Fractional Anisotropy (FA), Mean Diffusivity (MD), Axial Diffusivity (AD) results generated using our proposed model that leverages the attention mechanism.
- The proposed model efficiently learns spatial correlation in neighboring voxels, thereby improving the robustness of diffusion tensor imaging parameters.
- The proposed method speeds up the estimation of DTI parameters using sparse measurements.
- The proposed model exhibits the capability to quantitative diffusion parameters and identify early-stage Alzheimer's disease.
The following steps outline the data preprocessing workflow for DTI (Diffusion Tensor Imaging):
Apply the DTI model fit from the DIPY Python package Garyfallidis et al., 2014 to process each DWI (Diffusion-Weighted Imaging) image individually. This results in the computation of six diffusion components: Dxx, Dxy, Dyy, Dxz, Dyz, and Dzz of the diffusion tensor.
Select 100,000 voxels from each DWI image based on their fractional anisotropy (FA) score. Voxels with an FA score of zero are excluded, and the selection is uniformly distributed within the range (0, 1).
Generate 100,000 tuples of (input, ground-truth), where each tuple comprises:
- An input 5 × 5 × 5 voxel patch with 41 diffusion directions per voxel.
- A ground-truth 5 × 5 × 5 voxel patch with a 6 × 1 vector per voxel, representing the corresponding six diffusion components of the diffusion tensor.
Concatenate the diffusion signals of the neighboring 5 × 5 × 5 patches to construct a 125 × 41 matrix for each input voxel.
Combine the diffusion directions (3 × 41) with the diffusion signals to create a 128 × 41 matrix for each input voxel.
Zero-pad the input data for each tuple to form a 128 × 100 matrix. The corresponding ground truth is represented as a 125 × 6 matrix. This processed dataset is denoted as "ADNI − 41."
Perform Qball-based interpolation Tuch, 2004; Garyfallidis et al., 2014 on the 41-directional input data from the "ADNI − 41" dataset. This step yields 41-directional, 21-directional, and 5-directional diffusion signals, utilizing the DIPY package.
Combine the diffusion directions with the interpolated diffusion signals for each tuple. This results in input vectors of sizes 125 × 41, 125 × 21, and 125 × 5, along with a ground truth matrix of size 125 × 6.
Zero-pad the input data for each tuple to create a 128 × 100 matrix.
The resulting training datasets are named as follows:
- "41diff_proposedModel" generated 41 diffusion directions using the SwinTransformer model.
- "41diff_transformerDTI" generated 41 diffusion directions using the Davood Transformer model.
- "21diff_proposedModel" generated 21 diffusion directions using the SwinTransformer model.
- "21diff_transformerDTI" generated 21 diffusion directions using the Davood Transformer model.
- "5diff_proposedModel" generated 5 diffusion directions using the SwinTransformer model.
- "5diff_transformerDTI" generated 5 diffusion directions using the Davood Transformer model.
- The paper compares the diffusion direction of 41 using ground truth, proposed method, and LLS fitting Koay et al. (2006)
- The paper compares the diffusion direction of 21 using ground truth, pro-posed method, LLS fitting Koay et al. (2006), and Transformer-DTI Karimi and Gholipour (2022)
- The paper compares the diffusion direction of 5 using ground truth, proposed method, LLS fitting Koay et al. (2006), and Transformer-DTI Karimi and Gholipour (2022)
- Axial brain slice representing the Cingulum region, highlighted with p-value of two sample t-test(df=22). Green color: hypothesis testing tstat1 - Healthy CN > MCI, Red color: hypothesis testing tstat2 - Healthy CN < MCI.
- Coronal brain slice representing the Uncinate fasciculus region, highlighted with p-value of two sample t-test(df=22). Green color: hypothesis testing tstat1 - Healthy CN > MCI, Red color: hypothesis testing tstat2 - Healthy CN < MCI.
Name | Email ID |
---|---|
Abhishek Tiwari | at326@snu.edu.in |
Ananya Singhal | as146@snu.edu.in |
Dr. Saurabh J. Shigwan | saurabh.shigwan@snu.edu.in |
Dr. Rajeev Kumar Singh | Rajeev.kumar@snu.edu.in |