Learning to Generate Missing Pulse Sequence in MRI using Deep Convolution Neural Network Trained with Visual Turing...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuro-oncology

Background:

• Magnetic resonance imaging (MRI) utilizes multiple pulse sequences for comprehensive soft tissue visualization.
• Increased scan times and contrast allergies can limit the acquisition of all necessary MRI sequences.
• Synthesizing missing MRI sequences is crucial for applications like corrupted data correction and image enhancement.

Purpose of the Study:

• To develop and evaluate a deep convolution neural network (CNN) for synthesizing missing pulse sequences in brain MRI, particularly for tumor imaging.
• To address the limitations of acquiring complete MRI datasets due to scan time or patient-specific constraints.

Main Methods:

• A deep convolution neural network (CNN) with an encoder-decoder architecture was employed.
• The network was trained to minimize mean squared error (MSE) loss while incorporating an adversarial attack using a relativistic Visual Turing Test discriminator (rVTT).
• Experiments were conducted using the Brats2018 dataset for evaluation.

Main Results:

• The CNN successfully synthesized missing MRI pulse sequences for brain tumors.
• Quantitative metrics including MSE, Structural Similarity Measure (SSIM), and Peak Signal to Noise Ratio (PSNR) were used for evaluation.
• A Turing test conducted by radiologists and MR physicists achieved 76% accuracy, indicating superior performance compared to existing methods.
• The inference cost for image synthesis was determined to be 350.71 GFlops/voxel.

Conclusions:

• The developed CNN approach demonstrates superior performance in synthesizing missing MRI sequences for brain tumor imaging.
• This method offers a viable solution for scenarios with incomplete MRI data acquisition, enhancing diagnostic utility.
• The AI-driven synthesis of MRI sequences holds significant potential for improving clinical workflows and patient care.