Robust Polyp Detection and Diagnosis Through Compositional Prompt-Guided Diffusion Models.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Colorectal cancer (CRC) screening is vital for reducing mortality, with deep learning showing promise in polyp analysis.
• Current deep learning models struggle with generalization to diverse clinical data, especially out-of-distribution (OOD) data.
• Existing synthetic data generation methods, like diffusion models, often lack comprehensive clinical context.

Purpose of the Study:

• To develop an advanced diffusion model for generating clinically realistic synthetic colon polyp images.
• To improve the generalization and performance of deep learning models for polyp detection, classification, and segmentation.
• To address the limitations of current data augmentation and synthetic image generation techniques in medical imaging.

Main Methods:

• Proposed a Progressive Spectrum Diffusion Model (PSDM) integrating diverse clinical annotations (masks, bounding boxes, reports) into compositional prompts.
• Organized prompts into coarse and fine components to capture both broad structures and fine details.
• Augmented training datasets with PSDM-generated synthetic polyp images.

Main Results:

• PSDM-generated images enhance the performance of deep learning models for polyp detection, classification, and segmentation.
• Demonstrated significant improvements on the multi-center PolypGen dataset, increasing F1 score by 2.12% and mean average precision by 3.09%.
• Showcased superior performance in OOD scenarios, indicating enhanced model generalization.

Conclusions:

• The Progressive Spectrum Diffusion Model (PSDM) effectively generates clinically accurate synthetic polyp images by leveraging diverse annotations.
• PSDM-generated data augmentation significantly improves deep learning model performance and generalization for colorectal cancer screening.
• This approach offers a promising solution for overcoming data limitations and enhancing the robustness of AI in medical diagnostics.