Table 2.

Summary of the Architecture of the LoHiResGAN Model: A ResNet-based Generative Adversarial Network (GAN) for Efficient Translation of Low-Field to High-Field MRI Images.

Component	Details
Generator
Encoder (Downsampler)
Layers	8
Filter sizes	64, 128, 256, 512, 512, 512, 512, 512
Components	Convolutional layers, batch normalization, ReLU activation, residual blocks
Function	Reduces spatial dimensions of the input image, increase the number of feature maps
Decoder (Upsampler)
Layers	7
Filter sizes	512, 512, 512, 512, 256, 128, 64
Components	Transposed convolutional layers, batch normalization, ReLU activation, residual blocks, dropout layers (first three layers, dropout rate of 0.5)
Function	Increases spatial dimensions, decreases the number of feature maps
Output	Image of the same size as input, produced by a final transposed convolutional layer with a tanh activation function
Discriminator
Architecture	PatchGAN-based
Function	Classifies whether input image patches are real or generated
Inputs	Input image, target image (same size)
Downsampling Layers	3
Filter sizes	64, 128, 256
Components	Convolutional layers, batch normalization (except the first layer), LeakyReLU activation functions
Function	Reduces spatial dimensions of input, increases the number of feature maps
Output	$30\times 30$ map, each value corresponds to classification of a $70\times 70$ patch in the input image

LoHiResGAN is a GAN-based architecture that leverages ResNet components for efficient translation of low-field MRI images to high-field MRI images, which achieves improved image quality and structural preservation. Replacement of the U-Net downsampling and upsampling blocks with the ResNet counterparts in a modified U-Net generator can improve performance by leveraging the ResNet ability to capture long-range dependencies and preserve fine-grained details.