Unsupervised generative model for simulating post-operative double eyelid image

Area of Science:

• Medical Imaging
• Computer Vision
• Plastic Surgery Simulation

Background:

• Simulating double eyelid surgery outcomes is challenging, with current 3D models being complex and 2D methods yielding unnatural results.
• Existing 2D simulation techniques often require manual mask removal and struggle with realistic facial feature reconstruction.

Purpose of the Study:

• To develop an unsupervised generative model for simulating post-operative double eyelid surgery outcomes.
• To improve the realism and efficiency of double eyelid surgery simulations using 2D images.

Main Methods:

• A novel attention-class activation map module was integrated into a generative adversarial network (GAN).
• A dataset of pre- and post-operative 2D images was created for training.
• Adversarial consistency loss was adjusted to preserve source image features and eliminate masks.

Main Results:

• The proposed model successfully generated realistic double eyelid images.
• The attention module enhanced the generator's focus on relevant eyelid regions and improved the discriminator's accuracy.
• The method demonstrated superior performance compared to existing state-of-the-art techniques.

Conclusions:

• The unsupervised generative model with an attention mechanism offers a superior approach for double eyelid surgery simulation.
• This technique provides more natural and efficient simulations, overcoming limitations of previous methods.
• The model preserves essential facial features while effectively removing masks for improved visual outcomes.