TIST-Net: style transfer in dynamic contrast enhanced MRI using spatial and temporal information

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Deep Learning

Background:

• Training deep learning models for dynamic contrast-enhanced (DCE) MRI analysis is hindered by data scarcity and high acquisition costs.
• Variations in contrast enhancement within and between patients complicate model training.
• Style transfer offers a potential solution for generating synthetic medical images.

Purpose of the Study:

• To develop a style transfer method incorporating spatio-temporal information for DCE-MRI.
• To enable the addition or removal of contrast enhancement from existing DCE-MRI images.
• To create a method that can generate new DCE-MRI data to augment training datasets.

Main Methods:

• Proposed a temporal image-to-image style transfer network (TIST-Net) utilizing auto-encoders and convolutional long short-term memory networks.
• Employed deformable and adaptive convolutions for fine-grained control over content and style disentanglement in spatio-temporal data.
• Evaluated the method using structural similarity index measures (SSIM) and expert clinical evaluation.

Main Results:

• Achieved state-of-the-art performance on kidney, prostate, and uterus DCE-MRI datasets with high SSIM scores for both adding and removing contrast enhancement.
• Demonstrated superior performance compared to other methods in clinical evaluations, with experts consistently ranking TIST-Net higher.
• Successfully generated realistic DCE-MRI images by effectively disentangling content and style.

Conclusions:

• TIST-Net effectively generates novel DCE-MRI data from existing images by leveraging spatio-temporal information.
• The method shows significant potential for expanding limited training datasets for deep learning models.
• This technique could enhance the performance of models used for DCE-MRI image registration and segmentation.