CycleGAN for style transfer in X-ray angiography

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiovascular Simulation

Background:

• Data augmentation is crucial for training deep learning models in medical applications.
• Simulated medical images often lack the realism of real-world data.
• Enhancing the realism of simulated angiograms is essential for robust learning tasks.

Purpose of the Study:

• To generate realistic angiograms for data augmentation in medical learning tasks.
• To enhance the visual fidelity of simulated arterial images using deep learning.
• To improve the applicability of simulated data in training AI models for cardiovascular imaging.

Main Methods:

• Application of the CycleGAN deep neural network for style transfer.
• Transferring the visual characteristics of real angiograms to simulated arterial datasets.
• Utilizing an anatomically realistic cardiorespiratory simulator to generate baseline images.

Main Results:

• Achieved an average Structural Similarity Index (SSIM) of 0.948, indicating high fidelity in image translation.
• Vessel preservation was confirmed using the Dice coefficient, demonstrating effective maintenance of anatomical structures.
• Successful enhancement of simulated arterial images to mimic real angiographic appearances.

Conclusions:

• CycleGAN provides an effective method for enhancing artificial medical data, serving as an alternative to conventional data augmentation.
• The approach successfully generates realistic angiograms, respecting arterial physiology and X-ray patterns.
• This technique holds promise for improving deep learning applications in medical imaging, particularly for angiogram generation.