Synthesis of magnetic resonance images from computed tomography data using convolutional neural network with...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Deep Learning

Background:

• Magnetic resonance imaging (MRI) synthesis from computed tomography (CT) data is crucial for enhanced pathological visualization, particularly when MRI is unavailable.
• Convolutional neural networks (CNNs) offer a promising approach for synthesizing MRI from CT data.

Purpose of the Study:

• To develop and evaluate a novel deep CNN model for synthesizing T1- and T2-weighted MRI (T1WI and T2WI) from CT data.
• To compare the efficacy of different loss functions, including contextual loss, in MRI synthesis.

Main Methods:

• A dataset of 5,053 T1WI and 5,081 T2WI slices paired with CT images was curated and preprocessed.
• A double ResNet-U-Net (DRUNet) architecture was designed, integrating ResNet structures within a U-Net framework.
• The DRUNet model was trained and optimized using mean squared error (MSE), binary crossentropy (BCE), and contextual loss functions.

Main Results:

• DRUNet-101 with contextual loss significantly outperformed other models, achieving superior peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), and Tenengrad scores.
• Statistical analysis (P<0.001) confirmed the superiority of DRUNet-101 with contextual loss for both T1WI and T2WI synthesis.
• Visualizations confirmed the model's robustness and ability to generate synthetic MR images with high-frequency details.

Conclusions:

• DRUNet-101 with contextual loss demonstrates superior performance in synthesizing high-frequency information for MR images from CT scans.
• The proposed DRUNet model offers significant advantages over existing methods in terms of key image quality metrics.
• DRUNet-101 with contextual loss is recommended for synthesizing MR images from CT data in clinical settings.