CSG: A Context-Semantic Guided Diffusion Approach in De Novo Musculoskeletal Ultrasound Image Generation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Limited availability of diverse, unbiased medical data hinders AI development.
• Existing synthetic image generation methods struggle with semantic variability and context control.
• Need for advanced techniques to create representative medical image datasets.

Purpose of the Study:

• To introduce a scalable, dual-conditioned generative model (CSG) for enhanced control over synthetic medical image generation.
• To demonstrate CSG's capability in generating realistic musculoskeletal ultrasound images with pathological anomalies.
• To validate the quality and utility of CSG-generated synthetic images.

Main Methods:

• Developed a Context-Semantic Guidance (CSG) model for dual conditioning of image synthesis.
• Applied CSG to generate musculoskeletal ultrasound images, including pathological findings.
• Implemented a three-fold validation protocol including performance evaluation, similarity assessment, and Turing tests.
• Extended CSG to generate variations in anatomical geometries and textures.

Main Results:

• CSG-generated synthetic images improved semantic segmentation model performance.
• Synthetic images showed enhanced similarity to real ultrasound images compared to baseline methods.
• A Turing test confirmed that synthetic images were indistinguishable from real ones.
• CSG successfully generated clinically relevant pathological anomalies in MSK ultrasound images.

Conclusions:

• CSG offers comprehensive control over synthetic medical image generation, advancing realism and diversity.
• The model effectively synthesizes clinically valid pathological anomalies for musculoskeletal ultrasound.
• CSG-generated data is valuable for developing robust, unbiased AI models for disease detection and clinical decision support.