INTRODUCTION

Image quality transfer (IQT) is a computer vision-based approach to generate high-quality estimates of low-quality images. The usual methodology is to synthesise low-field training data by degrading high-field-acquired image data and to use this paired data to train either machine learning¹ or deep learning^2,3 models to generate a synthetic high-quality output.


SynthSR^2,4 is a software that generates a high-resolution T1-weighted image equivalent to any input scan. While SynthSR has demonstrated excellent results on synthetic low-field (<1T), data², it performs less well on image data acquired at ultra-low (<0.05T) fields (Fig 1). Moreover, SynthSR does not adapt well to abnormal morphology such as the presence of lesions (Fig 2). This may be related to the fact that SynthSR is designed to be contrast agnostic. This is achieved by manipulating separately the grey-scale voxel intensities of segmented brain structures to generate a low-field synthetic image, which is mapped to an authentic high-field T1-weighted image. The result is that the model learns to devalue information related to grey-scale values, and instead over-relies on features related to brain morphology. Because of this, brain lesion images may be problematic as these are generally characterised as a change in the expected grey-scale value in a region rather than a change in the large-scale morphology of the brain. As a result, SynthSR-enhanced images lesions appear to be "filled in" to reflect the general expected morphology in that region.

To avoid these problems, we generate synthetic low-resolution training images by degrading the high-resolution image as a whole (i.e. not treating structures separately). We also make use of K-space sub-sampling, gamma transforms and noise addition when degrading high-resolution images to generate synthetic low-resolution training images that qualitatively approximate ultra-low-field images. Finally, we adopt a ResUNet⁵ architecture (rather than a standard UNet^2,3) to learn the mapping between synthetic low-resolution and high-resolution images. When applied to authentic ultra-low-field acquired MR brain images, our results demonstrate an increase in image resolution, decrease in noise and an improvement in tissue contrast. We further demonstrate that our approach is adaptable to synthetic lesions.

METHODS

Data
Bias-field corrected T1-weighted images from the Human Connectome Project⁶ (HCP) were used as training data (n=100). Images were of dimensions [260x260x311], with voxels of side 0.7mm³. Test data were acquired on a Hyperfine system (0.064T). The T1-weighted images had dimensions [120x146x36], with voxels of side 1.6x1.6x5mm³. To evaluate model adaptability to abnormal morphology, we additionally generated a synthetic lesion dataset by manually drawing "lesions" (areas of consistent voxel intensity) on sample images from the HCP dataset. These were subsequently degraded to generate a synthetic low-resolution lesion image.

IQT Pipeline
In order to perform IQT we first degraded high-resolution image data to generate synthetic low-quality data. These paired data were then patched to smaller volumes and filtered to remove mean-zero patches before being used to train a ResUNet model.

Image Preprocessing
The preprocessing pipeline is outlined in Fig 3. Synthetic low-resolution data was generated by first skull-stripping image data, standardising and subsequently cubing image dimensions to [256x256x311]. Thereafter image resolution was reduced by a factor [2x2x8] to reflect reduced dimensions of ultra-low field images. The images were FFT transformed into K-space and sub-sampled. Noise was added before transformation back to image space and applying Gaussian kernel for blurring. The images were then restored to original dimensions using cubic spline interpolation, and a mask applied to remove noise from non-brain voxels. The hyperparameters controlling the degree of K-space subsampling, noise and blurring were generated probabilistically to provide a diverse range of training images of varying quality.

For testing, the resolution of Hyperfine images was increased using cubic spline interpolation. The images were then patched and the trained model applied to each patch. All patches were concatenated to generate a final IQT image.

Training
Training images were patched into volumetric slices of side [256x256x3]. Patches were filtered to exclude those with mean zero. A 4-layer ResUNet architecture⁴ was implemented in PyTorch and run on Google Colab GPU. Due to computational limitations, the number of filters in each layer was set at [32, 64, 120, 180]. Early stopping criteria was used to end training. The Adam optimizer was used to minimize the L1 loss which has been shown to produce visually sharper boundaries and facilitate training by escaping local minima.⁷

RESULTS

Results of the trained model applied to authentic ultra-low-field image data are reported in Fig. 4. They show improved tissue contrast, decreased noise and increased resolution. However, the use of thin volumetric patches leads to image artefacts and decreased tissue detail in out-of-plane image slices. This may be addressed in future studies by adopting cubic image patches for training.

Qualitatively, the trained model performed well when applied to synthetic low-resolution lesion data (Fig 5), with the resulting IQT’d volume retaining the synthetic lesion.

DISCUSSION & CONCLUSION

Image synthesis for training can play a key role in determining the type of features learned by models designed for IQT. Improvements in synthetic low-resolution image generation and model architecture can significantly benefit the performance of IQT on images acquired at ultra-low field strengths. The results presented here should be validated on external datasets.

Acknowledgements

No acknowledgement found.