Towards image registration of dynamic contrast enhanced MRI using deep learning

Dynamic contrast enhanced (DCE)-MRI is a quantitative imaging technique used to monitor microvascular perfusion. It involves rapidly acquiring multiple T1-weighted images over a few minutes whilst injecting a contrast agent, which causes a rapid increase in image intensity. Each type of tissue exhibits unique intensity enhancement characteristics over time. Typically, DCE-MRI data undergoes quantitative analysis, often employing tracer kinetics modelling, which necessitates manually selecting voxels in each frame that represent the tissue of interest. However, due to motion between frames, these selected voxels may not consistently represent the target tissue over the temporal dimension. To mitigate this issue, image registration is used to align images, reducing the impact of motion and allowing for more accurate analysis. In this thesis, I began by exploring which reference image should be used for image registration. This decision is crucial because if an image is chosen poorly, the registration accuracy will be low. Several approaches were identified and tested, with findings showing that using an image with contrast enhancement was the best approach. Additionally, I investigated how contrast enhancement in DCE-MRI affects the performance of image registration and segmentation. The effects of contrast enhancement when using deep learning in DCE-MRI image processing tasks has not been previously shown. In this work, images were split into groups based on the amount of contrast enhancement in the image. From this, datasets were created and used to train and test models. The results showed that strategically using available data (pre-training and fine tuning using different splits of contrast enhancement) leads to better-performing, robust models. However, achieving this effectively requires a diverse dataset. This led to the second piece of work, which explored image synthesis to expand existing DCE-MRI datasets. This is challenging because the generated images must be realistic, containing key structures and varying levels of contrast enhancement. I proposed a style transfer method that utilised temporal information to generate new images that structurally resembled an input content image whilst exhibiting different amounts of contrast enhancement. Additionally, I introduced a new metric to evaluate the quality of the generated images based on their style and content. These images were also evaluated qualitatively by MRI experts. The new method successfully generated synthetic DCE-MR images, which can supplement existing datasets for training deep learning models. Using these expanded datasets, more powerful architectures can be used for image registration, such as transformer-based models. In this work, I used a hybrid model composed of shifted window (Swin) transformers and convolutional long short-term memory (LSTM) models. LSTMs were used to model the temporal dimension, whilst the Swin transformers captured the relationship between pairs of images. This model outperformed popular deep learning image registration methods, likely due to transformers’ increased receptive field compared to traditional convolutional kernels.

Grantor dc:publisher

The University of Edinburgh

Year dc:date.issued

2025

Author dc:creator

• Tattersall, Adam

Advisors dc:contributor.advisor

• Kershaw, Lucy
• Semple, Scott

• Dynamic contrast enhanced-MRI
• DCE-MRI
• image alignment
• changes in brightness
• segmentation of different tissues
• deep learning image processing
• synthetic DCE-MR images
• AI training data

Language dc:language.iso

en

Identifier URI

http://dx.doi.org/10.7488/era/5656

OAI identifier oai:identifier

oai:era.ed.ac.uk:1842/43113