Deep learning based apparent diffusion coefficient map generation from multi-parametric MR images for patients with...

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Medicine
• Biomedical Engineering

Background:

• Apparent diffusion coefficient (ADC) maps from diffusion-weighted magnetic resonance imaging (DWI MRI) offer functional tissue insights.
• DWI MRI is time-consuming and prone to artifacts, compromising ADC map accuracy.
• Synthesizing ADC maps from multi-parametric MRI could overcome these limitations.

Purpose of the Study:

• To develop a deep learning framework for synthesizing ADC maps from multi-parametric MRI.
• To address the challenges of time consumption and artifacts associated with traditional DWI MRI.

Main Methods:

• Proposed the multiparametric residual vision transformer (MPR-ViT) model, integrating Vision Transformer (ViT) and convolutional operators.
• Utilized T1-weighted (T1w) and T2-fluid attenuated inversion recovery (T2-FLAIR) images from 501 glioma cases.
• Evaluated model performance against Vision Convolutional Transformer (VCT) and residual vision transformer (ResViT) using PSNR, SSIM, and MSE metrics.

Main Results:

• The MPR-ViT model achieved a PSNR of 31.0 ± 2.1, SSIM of 0.950 ± 0.015, and MSE of 0.009 ± 0.0005 using T1w + T2-FLAIR MRI inputs.
• Ablation studies demonstrated the performance impact of individual input sequences.
• Qualitative and quantitative analyses confirmed the MPR-ViT model's favorable performance compared to ground truth data.

Conclusions:

• High-quality ADC maps can be successfully synthesized from structural MRI using the MPR-ViT model.
• The predicted ADC maps exhibit superior conformity to ground truth volumes compared to VCT and ResViT.
• Synthetic ADC maps are valuable for disease diagnosis and intervention, especially when conventional ADC maps are compromised or unavailable.