SinGAN-Seg

SinGAN-Seg generates synthetic medical images and corresponding segmentation masks from a single annotated image to augment datasets for medical image segmentation and mitigate data scarcity and privacy constraints.

Key Features:

• Single-Image Training: Trains using only one image and its corresponding ground-truth mask, reducing dependency on large annotated datasets.
• Synthetic Data Generation: Produces synthetic medical images together with segmentation masks that can substitute for real data when sharing is restricted.
• Style Transfer Technique: Applies style transfer to enhance the realism and diagnostic appearance of generated images.
• Performance Evaluation: Assesses outputs using qualitative and quantitative comparisons between real and synthetic datasets to benchmark generation quality.
• Model Training Enhancement: Combines synthetic and real data to train segmentation models such as UNet++, enabling performance comparable to training on larger real datasets.

Scientific Applications:

• Segmentation model augmentation: Augments training data for medical image segmentation tasks to improve model robustness under limited-data conditions.
• Rare abnormality representation: Generates examples of scarce or hard-to-obtain abnormalities to support training and evaluation of segmentation algorithms.
• Oncology, radiology, and pathology: Supplies synthetic annotated images for segmentation tasks across oncology, radiology, and pathology imaging studies.

Methodology:

Train on a single image and its ground-truth mask to generate synthetic images and corresponding masks, apply a style transfer technique to increase realism, and evaluate outputs via qualitative and quantitative comparisons and downstream segmentation performance (e.g., UNet++ trained with combined real and synthetic data).