HiDiff: Hybrid Diffusion Framework for Medical Image Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning (DL) models for medical image segmentation are primarily discriminative, learning input-image-to-mask mappings.
• These discriminative models often overlook underlying data distributions and class characteristics, leading to unstable feature representations.

Purpose of the Study:

• To enhance medical image segmentation by integrating generative model knowledge into discriminative approaches.
• To introduce HiDiff, a novel hybrid diffusion framework that synergizes discriminative and generative diffusion models for improved segmentation.

Main Methods:

• HiDiff employs a two-component architecture: a discriminative segmentor providing an initial mask and a novel binary Bernoulli diffusion model (BBDM) for refinement.
• The segmentor and BBDM are trained in an alternate-collaborative manner to mutually enhance their performance.
• The BBDM models the underlying data distribution to effectively refine segmentation masks.

Main Results:

• HiDiff demonstrated superior performance across diverse medical image segmentation tasks (abdomen organs, brain tumors, polyps, retinal vessels) and modalities.
• The framework outperformed existing state-of-the-art transformer- and diffusion-based segmentation algorithms.
• HiDiff showed particular strength in segmenting small objects and generalizing to new, unseen datasets.

Conclusions:

• The proposed HiDiff framework effectively combines discriminative and generative approaches for robust medical image segmentation.
• HiDiff offers a significant advancement in medical image segmentation, particularly for challenging cases involving small structures and cross-dataset generalization.
• The hybrid approach provides a promising direction for future developments in medical image analysis.