Few-shot biomedical image segmentation using diffusion models: Beyond image generation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Medical image segmentation demands extensive annotated data, which is costly and time-consuming.
• Generative models offer a potential solution for few-shot image segmentation tasks.

Purpose of the Study:

• To leverage generative models for efficient few-shot medical image segmentation.
• To evaluate the effectiveness of denoising diffusion probabilistic models (DDPMs) as feature extractors for segmentation.

Main Methods:

• A DDPM was trained on a large dataset of pelvis radiographs to generate synthetic images.
• Features were extracted from real images using the pre-trained DDPM at multiple timesteps.
• A U-Net model was trained using these extracted features for landmark segmentation.

Main Results:

• The generated images were validated by experts and objective metrics (FID=7.2, IS=210).
• The U-Net model trained with DDPM features achieved high Dice Similarity Coefficients (DSC) for segmenting obturator foramen (0.90), greater trochanter (0.84), and lesser trochanter (0.61).
• Performance was significantly better than a U-Net trained without DDPM features.

Conclusions:

• DDPMs can be effectively utilized as feature extractors in medical image analysis.
• This method facilitates accurate medical image segmentation even with limited annotated data.