Utilizing conditional generative adversarial network to generate head MRA based on nonvascular sequences: comparative...

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence in Medicine

Background:

• Magnetic Resonance Angiography (MRA) is crucial for vascular imaging.
• Acquiring MRA often requires specific nonvascular sequences, increasing scan time and complexity.
• Deep learning offers potential for synthesizing MRA from readily available nonvascular MRI sequences.

Purpose of the Study:

• To develop a deep learning model for synthesizing head MRA from T1W, T2W, and FLAIR sequences.
• To evaluate the diagnostic performance and image quality of synthetic MRA (syn-MRA) compared to real MRA.

Main Methods:

• A retrospective study of geriatric inpatients with multimodal MRI data.
• Development of single-modality (T1W, T2W, FLAIR) and multi-modality (MIX) conditional generative adversarial network (cGAN) models.
• Quantitative image quality assessment and qualitative evaluation by radiologists for visual quality, diagnostic confidence, and structured reports.

Main Results:

• The multi-modality MIX model outperformed single-modality models in quantitative metrics.
• Syn-MRA demonstrated significantly higher signal-to-noise and contrast-to-noise ratios than real MRA.
• The MIX model achieved comparable overall image quality, vessel sharpness, and diagnostic confidence to real MRA, with 92.2% diagnostic accuracy.
• Venous contamination was more pronounced in syn-MRA.

Conclusions:

• A deep learning-based multi-modality model (MIX) can effectively synthesize head MRA from nonvascular sequences.
• The developed syn-MRA exhibits diagnostic performance comparable to real MRA.
• This approach holds potential for efficient and accessible vascular imaging in clinical practice.