A generic plug & play diffusion-based denosing module for medical image segmentation.

Area of Science:

• Computer Vision
• Medical Imaging Analysis
• Deep Learning

Background:

• Medical image segmentation is crucial but challenged by small datasets, noise, and artifacts.
• Diffusion models show promise in image generation and computer vision tasks.

Purpose of the Study:

• To introduce a novel feature map denoising (FMD) module for enhancing medical image segmentation.
• To integrate the FMD module into existing segmentation networks for seamless end-to-end training.

Main Methods:

• Developed a plug-and-play feature map denoising (FMD) module based on diffusion models.
• Integrated the FMD module into UNet, UNeXt, TransUNet, and IB-TransUNet models.
• Evaluated performance across four diverse datasets.

Main Results:

• The FMD module significantly improved the performance of all tested segmentation models.
• Enhanced segmentation accuracy was observed, especially for small lesion areas and minor organs.
• The FMD module demonstrated a positive impact on overall model performance.

Conclusions:

• The proposed FMD module effectively refines feature maps, leading to more accurate medical image segmentation.
• The plug-and-play nature allows flexible integration, offering a versatile solution for segmentation challenges.
• This approach shows particular benefit in segmenting challenging small structures within medical images.