Diffusion semantic segmentation model: A generative model for medical image segmentation based on joint distribution.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Current medical image segmentation primarily uses discriminative models based on conditional distributions p(class|feature).
• These models neglect the underlying data distribution p(feature|class), leading to unstable feature spaces and imprecise anatomical boundary delineation.

Purpose of the Study:

• To reformulate semantic segmentation as a distribution alignment problem for enhanced medical image segmentation.
• To minimize the discrepancy between model predictions and ground truth by modeling the joint data distribution.

Main Methods:

• Propose a novel Denoising Semantic Segmentation Model (DSSM) architecture.
• Learn classification boundaries in pixel feature space and model joint distributions in latent feature space.
• Utilize Bayesian posterior probabilities for optimizing probability maps and a Feature Fusion Module (FFM) for guiding inference.

Main Results:

• DSSM demonstrates superior performance over state-of-the-art discriminative models across diverse modalities (MRI, X-ray, skin lesions).
• Achieved high Dice coefficients, e.g., 0.9647 in X-ray segmentation and 0.9421 in PH2 dataset skin lesion segmentation.
• Further validated by metrics like HD95, mIoU, Precision, and Recall, confirming enhanced segmentation accuracy.

Conclusions:

• The proposed methodology stabilizes the learned feature space by capturing latent feature distribution information.
• DSSM significantly outperforms traditional discriminative segmentation methods on multi-modal datasets.