Learning ADC maps from accelerated radial k-space diffusion-weighted MRI in mice using a deep CNN-transformer model.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Diffusion MRI

Background:

• Diffusion-weighted spin-echo (DW-SE) imaging is crucial for generating ADC maps.
• Accelerated acquisition methods are needed to reduce scan times.
• Radially sampled DW-SE (Rad-DW-SE) offers potential for acceleration.

Purpose of the Study:

• To accelerate the Rad-DW-SE acquisition method for high-quality ADC map generation.
• To develop and validate a deep learning approach for accelerated ADC mapping.

Main Methods:

• A deep learning model integrating convolutional neural networks (CNNs) and vision transformers was developed.
• The model generates ADC maps from accelerated DWI data, regularized by a monoexponential ADC model fitting term.
• Training and evaluation were performed on DWI data from 147 and 36 mice, respectively, using 4x and 8x acceleration factors.

Main Results:

• The proposed deep learning model produced higher quality ADC maps compared to alternative methods.
• Performance was evaluated on whole images and specific regions of interest (tumors, kidneys, muscles).
• Ablation studies confirmed the model's effectiveness in generating accurate ADC maps from accelerated data.

Conclusions:

• The deep learning method effectively computes accurate ADC maps from accelerated Rad-DW-SE DWI data.
• Integration of CNNs and transformers enhances ADC map quality and accuracy.
• This approach facilitates faster and more precise diffusion MRI analysis.