Pseudo-CT image generation from mDixon MRI images using fully convolutional neural networks

Generating pseudo-CT images from MRI provides electron density maps for radiation therapy planning and saves additional CT scans. Fully convolutional neural networks were proposed for pseudo-CT generation. We investigated the influence of architectures and hyperparameters on the quality of the pseudo-CT images. We used fully convolutional neural networks to transform between registered MRI and CT volumes of the pelvic region: two UNet variants using transposed convolutions or bilinear upsampling, LinkNet using residual blocks and strided convolutions for downsampling, and we designed transnet to maintain tensor spatial dimensions equal to the image’s size. Different architectures revealed similar error metrics, although pseudo-CTs differ visually. Comparison of LinkNet and UNet showed that downsampling does not affect translation. Replacing transposed convolutions with bilinear upsampling improved the pseudo-CTs’ sharpness. Translation quality quickly saturates with the number of convolution layers; increasing the number of layers from 4 to 19 decreases the MAE from 44HU to 37HU. Varying the number of feature maps showed that good translation quality can be achieved with networks that are substantially narrower than those previously published. Generally, the pseudo-CT have MAE lower than 45HU, computed inside of the true CT’s body shape.