Multi-Conditional Diffusion Framework with Texture Constraints for Clinically-Reliable Lesion Synthesis in Virtual...

Area of Science:

• Artificial Intelligence in Medical Imaging
• Computational Pathology
• Medical Image Synthesis

Background:

• Clinical translation of AI in medical imaging is hindered by limited annotated data, imbalanced datasets, and privacy concerns.
• Existing synthetic data generation methods struggle to maintain pathological heterogeneity and structural authenticity.

Purpose of the Study:

• To develop a novel framework for synthesizing clinically reliable lesions in CT images.
• To address limitations in data availability, pathological distribution, and privacy for virtual imaging trials.

Main Methods:

• Proposed a Multi-conditional Diffusion framework with Texture Constraints (MDTC).
• Integrated anatomical mask guidance and a Gray-Level Co-occurrence Matrix (GLCM) texture classifier into the diffusion network.
• Employed dual-constraint mechanisms to enforce structural fidelity and pathological heterogeneity.

Main Results:

• MDTC demonstrated superior texture fidelity, improved Peak Signal-to-Noise Ratio (PSNR), and optimized Fréchet Inception Distance (FID) compared to traditional models.
• Downstream classifiers trained on MDTC-generated data maintained high classification performance.
• Data augmentation with synthetic rare pulmonary nodules significantly enhanced classification efficacy.

Conclusions:

• MDTC offers a new paradigm for generating diagnostically meaningful synthetic data, overcoming critical bottlenecks in virtual imaging trials.
• The framework preserves structural authenticity and mitigates texture homogenization, enabling robust diagnostic model training.
• MDTC provides a practical, privacy-preserving solution for expanding virtual imaging trials to underrepresented diseases.