Image Distortion Correction for Diffusion MR Imaging Using a Transformer-based U-Net.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology
• Neuroimaging Analysis

Background:

• Diffusion-weighted imaging (DWI) is crucial for assessing brain microstructure.
• Image distortions in DWI, particularly at high b-values, can impede accurate analysis.
• Existing correction methods, like U-Net-based approaches, have limitations.

Purpose of the Study:

• To evaluate TransDisCo, a Transformer-based U-Net, for clinical DWI distortion correction.
• To compare TransDisCo's performance against a U-Net-based method (Synb0).
• To assess performance across various DWI techniques, including high b-value DWI, DTI, and DKI, on clinical data with pathologies.

Main Methods:

• Utilized a clinical DWI dataset from patients with brain tumors and cerebrovascular disorders.
• Compared TransDisCo with Synb0, training both on T1-weighted and uncorrected DWI images.
• Conducted an ablation study replacing Transformer blocks with convolutional blocks to assess the self-attention mechanism's contribution.
• Evaluated corrected images and derived maps using Peak SNR, SSIM, MI, and Dice coefficients.

Main Results:

• TransDisCo and Synb0 successfully corrected distortions in clinical DWI, DTI, and DKI data.
• TransDisCo significantly outperformed Synb0 in quantitative metrics, preserving anatomical details and reducing artifacts, especially in pathological cases.
• The ablation study confirmed that TransDisCo's self-attention mechanism enhanced performance over a CNN-based variant.

Conclusions:

• TransDisCo offers a promising advancement over conventional CNN-based and U-Net-based methods for DWI distortion correction.
• The Transformer-based approach effectively handles diverse clinical pathologies and imaging conditions.
• TransDisCo has the potential to enhance the precision and reliability of diffusion MRI applications.