Colorectal polyp segmentation with denoising diffusion probabilistic models.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Early detection of polyps is critical for reducing colorectal cancer (CRC) incidence.
• Accurate polyp segmentation is essential for effective clinical CRC prevention strategies.
• Current segmentation techniques require improvement in efficiency and accuracy.

Purpose of the Study:

• To develop an efficient and accurate end-to-end polyp segmentation technique using a diffusion model.
• To formulate polyp segmentation as a mask generation process leveraging image priors.
• To enhance segmentation performance through a majority vote strategy on multiple predictions.

Main Methods:

• An end-to-end training approach employing a diffusion model for polyp segmentation.
• Images are treated as priors, with segmentation defined as a mask generation task.
• A majority vote strategy is applied to multiple predictions generated during the sampling process.

Main Results:

• The proposed diffusion model achieved high performance on multiple datasets, including mDice scores of 0.934 (Kvasir-SEG) and 0.967 (CVC-ClinicDB).
• Cross-validation demonstrated strong generalization capability, outperforming existing state-of-the-art models.
• The method significantly improves segmentation accuracy for polyp detection.

Conclusions:

• The proposed diffusion model offers a promising approach for accurate and efficient polyp segmentation.
• This technique has the potential to significantly advance early detection and prevention of colorectal cancer.
• The model's strong generalization capability suggests broad applicability in clinical settings.