Deep learning based MRI contrast synthesis using full volume prediction using full volume prediction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Magnetic Resonance Imaging (MRI) generates multiple image contrasts offering complementary information.
• Acquiring multiple MRI contrasts is often limited by time and patient comfort.
• Existing contrast synthesis methods can lead to inconsistencies in 3D volumes.

Purpose of the Study:

• To develop a 3D convolutional neural network (CNN) for synthesizing T2 and FLAIR MRI contrasts from a single T1 volume.
• To overcome the slice-wise intensity inconsistencies of previous 2D-based contrast synthesis methods.
• To improve the accuracy and coherence of synthesized MRI contrasts for enhanced analysis.

Main Methods:

• A 3D UNet-variant convolutional neural network (CNN) was employed to process entire MRI volumes.
• Spatial-to-depth and reconstruction layers were introduced to manage memory demands for full-volume processing.
• The network was trained to generate T2 and FLAIR images from T1 source volumes.

Main Results:

• The proposed 3D CNN generated synthesized MRI volumes with enhanced slice-to-slice coherence.
• The method demonstrated improved accuracy in contrast synthesis due to 3D context awareness.
• Validation using a segmentation task confirmed the practical utility of the synthesized contrasts.

Conclusions:

• The 3D CNN approach effectively synthesizes consistent and accurate MRI contrasts (T2, FLAIR) from T1 volumes.
• This method addresses limitations of slice-based synthesis, offering a more robust solution for multi-contrast MRI.
• The synthesized contrasts are valuable for downstream applications like MRI segmentation, enhancing diagnostic capabilities.