Deep Learning for Breast MRI Style Transfer with Limited Training Data

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Medical image datasets often exhibit style variations due to different acquisition protocols and scanner models.
• These style discrepancies can hinder the performance of machine learning models in downstream tasks.
• Existing style transfer methods may require extensive training data or complex optimization procedures.

Purpose of the Study:

• To introduce StyleMapper, a novel medical image style transfer method.
• To enable efficient and arbitrary style transfer, even to unseen styles, with limited training data.
• To facilitate the creation of unified medical image datasets for improved downstream task performance.

Main Methods:

• StyleMapper employs a novel approach trained on diverse simulated styles, enhancing computational efficiency.
• The model disentangles image content from style, allowing style modification by replacing style encodings.
• Arbitrary style transfer is achieved by using a single target style image without additional optimization.

Main Results:

• Experimental results on breast magnetic resonance images demonstrate the effectiveness of StyleMapper for style transfer.
• The method successfully transfers medical scans to unseen styles.
• The approach allows for the alignment of images from different scanners into a unified style.

Conclusions:

• StyleMapper offers an efficient and versatile solution for medical image style transfer.
• The method's ability to handle unseen styles and limited data makes it valuable for real-world medical imaging applications.
• Unified datasets generated by StyleMapper can enhance the training and performance of various downstream medical image analysis tasks.