Exploring the Acceleration Limits of Deep Learning Variational Network-based Two-dimensional Brain MRI.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Deep learning models are increasingly used for Magnetic Resonance Imaging (MRI) reconstruction.
• Accelerated MRI acquisition techniques aim to reduce scan times, improving patient comfort and throughput.
• Understanding the limits of deep learning reconstruction at high acceleration is crucial for clinical applications.

Purpose of the Study:

• To explore the performance limits of deep learning-based brain MRI reconstruction.
• To identify optimal acceleration ranges for general-purpose diagnostic imaging and potential screening protocols.
• To evaluate the model's performance on both in-domain and out-of-domain data.

Main Methods:

• A deep learning model was trained on 5847 brain MRI images.
• Performance was assessed across various acceleration factors (up to 100×) using the fastMRI dataset.
• Radiologists evaluated reconstruction quality for diagnostic and screening use cases.
• A Monte Carlo procedure estimated reconstruction error.

Main Results:

• Radiologists deemed 100% of reconstructions sufficient for general-purpose imaging up to 4× acceleration.
• 94% of reconstructions were considered adequate for screening protocols up to 14× acceleration.
• The model demonstrated robust performance even at 100× acceleration on out-of-distribution data.

Conclusions:

• Deep learning-based MRI reconstruction allows for significantly higher acceleration factors for screening compared to diagnostic imaging.
• The findings suggest potential for accelerated MRI protocols in both clinical diagnosis and large-scale screening.
• Further research can optimize deep learning reconstruction for ultra-high acceleration in brain MRI.