Using diffusion models to generate synthetic labeled data for medical image segmentation.

Area of Science:

• Medical imaging
• Machine learning
• Computer vision

Background:

• Machine learning is increasingly applied to medical image analysis.
• Limited availability of high-quality, annotated medical data due to privacy and cost constraints hinders progress.
• Synthetic data generation offers a potential solution to data scarcity in medical AI.

Purpose of the Study:

• To design and evaluate a pipeline for generating synthetic labeled polyp images.
• To augment medical image segmentation models and address data scarcity.
• To reduce the reliance on manual annotation for training machine learning models.

Main Methods:

• Diffusion models were trained on the HyperKvasir dataset of gastrointestinal polyp images.
• Generated images were evaluated using qualitative expert review, Fréchet Inception Distance (FID), and Multi-Scale Structural Similarity (MS-SSIM).
• Segmentation models were trained with synthetic data and evaluated using Dice Score (DS) and Intersection over Union (IoU).

Main Results:

• The pipeline generated synthetic polyp images comparable to real images, as indicated by FID scores.
• Segmentation models trained with synthetic data showed improved performance over Generative Adversarial Network (GAN) methods.
• Performance gains were observed even with partial training on synthetic data, with demonstrated transferability across datasets.

Conclusions:

• The developed pipeline effectively produces realistic synthetic image-mask pairs for medical imaging tasks.
• The synthetic data generation approach can significantly reduce the need for manual data annotation.
• Training segmentation models fully or partially on the generated synthetic data enhances performance, as evidenced by Dice and IoU scores.