ColoDiff: Integrating Dynamic Consistency With Content Awareness for Colonoscopy Video Generation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Biology

Background:

• Colonoscopy video generation is vital for diagnosing intestinal diseases, especially with limited data.
• Challenges include maintaining temporal consistency, controlling clinical attributes, and handling irregular intestinal structures and diverse imaging modalities.

Purpose of the Study:

• To develop ColoDiff, a diffusion-based framework for generating dynamic-consistent and content-aware colonoscopy videos.
• To address data scarcity in medical imaging and assist in clinical analysis.

Main Methods:

• ColoDiff utilizes a TimeStream module for inter-frame temporal dependency modeling via cross-frame tokenization.
• A Content-Aware module with noise-injected embeddings and learnable prototypes enables precise control over clinical attributes.
• A non-Markovian sampling strategy achieves over 90% reduction in generation steps for real-time performance.

Main Results:

• ColoDiff successfully generates colonoscopy videos with smooth transitions and rich dynamics.
• The framework produces customized content for various diagnostic tasks (e.g., colitis, polyps, adenomas).
• Incorporating synthetic videos into training improved diagnosis accuracy by 7.1%.

Conclusions:

• ColoDiff demonstrates effective controllable colonoscopy video generation.
• Synthetic videos generated by ColoDiff can complement authentic data and mitigate scarcity in clinical settings.
• The framework shows potential for enhancing medical training and diagnostic tools.