Guided image generation for improved surgical image segmentation.

Area of Science:

• Medical Imaging
• Machine Learning
• Computer-Aided Surgery

Background:

• Limited large, annotated datasets hinder accurate machine learning (ML) model development in surgery.
• Generative models create synthetic data but face challenges in surgical domains like anatomical diversity.
• Existing models may produce overfitted, unrealistic, or cartoonish surgical images.

Purpose of the Study:

• Introduce Surgery-GAN, a novel generative model for synthetic surgical image creation from segmentation maps.
• Enhance synthetic surgical image quality and diversity compared to existing generative models.
• Evaluate the impact of synthetic data augmentation on ML segmentation model performance.

Main Methods:

• Developed Surgery-GAN, a generative adversarial network incorporating channel- and pixel-level normalization.
• Trained and tested Surgery-GAN on cholecystectomy, partial nephrectomy, and radical prostatectomy datasets.
• Utilized generated synthetic data to train five distinct ML segmentation models alongside real data.

Main Results:

• Surgery-GAN produced novel, realistic, and diverse surgical images across three surgical datasets.
• Synthetic images consistently improved mean segmentation performance when combined with real data.
• Significant performance boosts observed, particularly for under-represented classes (up to 61.6% increase).

Conclusions:

• Surgery-GAN effectively generates high-quality, diverse synthetic surgical images.
• Augmenting real datasets with Surgery-GAN-generated images enhances ML segmentation model accuracy.
• The model shows particular promise for improving segmentation of rare anatomical structures in surgical datasets.