Accurate multi-b-value DWI generation using two-stage deep learning: multicenter study.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Diffusion-weighted imaging (DWI) is crucial for quantitative analysis but faces limitations in acquisition.
• Synthesizing multi-b-value DWI and restoring apparent diffusion coefficient (ADC) maps accurately is challenging.
• Existing methods struggle with high-fidelity synthesis across diverse anatomical regions.

Purpose of the Study:

• To develop and validate a two-stage deep learning framework (DC²Anet-MineGAN) for multi-organ, multi-b-value DWI synthesis.
• To achieve accurate ADC restoration from synthesized DWI data.
• To address real-world limitations in clinical DWI acquisition.

Main Methods:

• A retrospective study utilized 50,000 DWI images from three hospitals and TCIA database across five anatomical regions and various b-values.
• A two-stage model, DC²Anet for low-to-high b-value synthesis and MineGAN for interpolation, was employed.
• Performance was evaluated using quantitative metrics (MSE, MAE, PSNR, SSIM) and radiologist Likert ratings with ICC.

Main Results:

• Synthetic ADC values closely matched ground truth across all regions (mean difference < 0.02; p > 0.05).
• High image quality was confirmed with SSIM > 0.81 and PSNR > 74 for all b-values.
• Radiologists rated 75% and 50% of synthetic images at unseen b-values as excellent, with ICC exceeding 0.92.

Conclusions:

• The DC²Anet-MineGAN framework enables accurate, high-quality DWI synthesis and ADC reproduction.
• The model overcomes clinical DWI limitations, supporting reliable quantitative imaging across multiple b-values and anatomical regions.
• Further multi-center clinical validation is recommended to address potential hallucination or distortion.