AnatoDiff: Síntesis de radiografías anatómicamente veraces con imágenes de entrenamiento limitadas

Área de la Ciencia:

• Medical Imaging
• Artificial Intelligence
• Radiography

Sus antecedentes:

• Diffusion models can synthesize realistic radiographic images but require extensive training data (often >10,000 images).
• Pre-training on natural images is insufficient for generating anatomically accurate medical images due to domain differences.
• Limited data availability hinders the application of current models in specialized medical conditions.

Objetivo del estudio:

• To develop AnatoDiff, a diffusion model capable of synthesizing high-quality X-ray images with accurate anatomical shapes using limited training data (500-1,000 images).
• To enable effective medical image synthesis for specialized conditions with scarce datasets.

Principales métodos:

• AnatoDiff integrates a Shape Prototype Module and an Anatomical Fidelity loss function for targeted supervision.
• The model was validated on three diverse datasets: Neonatal Abdomen, Adult Chest, and Humerus.
• Performance was compared against state-of-the-art few-shot and data-limited synthesis methods.

Principales resultados:

• AnatoDiff achieved significant improvements in image synthesis quality, with average gains of 14.9% in Fréchet Inception Distance, 9.7% in Improved Precision, and 2.3% in Improved Recall.
• Generated images demonstrated consistent anatomical accuracy, outperforming existing models.
• A classifier trained on AnatoDiff images showed a 2.1%-5.3% increase in F1-score compared to training on state-of-the-art diffusion images.
• Medical professionals found AnatoDiff images difficult to distinguish from real radiographs.

Conclusiones:

• AnatoDiff effectively synthesizes high-quality, anatomically accurate X-ray images using substantially reduced training data.
• The model's targeted supervision mechanisms enable high performance in data-limited scenarios.
• AnatoDiff holds promise for advancing medical imaging research and applications, particularly in rare diseases or specialized areas.