AnatoDiff: Synthesizing Anatomically Truthful Radiographs With Limited Training Images

Rapid advancements in diffusion models have enabled synthesis of realistic and anonymized imagery in radiography. However, due to their complexity, these models typically require large training volumes, often exceeding 10,000 images. Pre-training on natural images can partly mitigate this issue, but often fails to generate anatomically accurate shapes due to the significant domain gap. This prohibits applications in specialized medical conditions with limited data. We propose AnatoDiff, a diffusion model synthesizing high-quality X-Ray images with accurate anatomical shapes using only 500 to 1,000 training samples. AnatoDiff incorporates a Shape Prototype Module and Anatomical Fidelity loss, allowing for smaller training volumes through targeted supervision. We extensively validate AnatoDiff across three open-source datasets from distinct anatomical regions: Neonatal Abdomen (1,000 images); Adult Chest (500 images); and Humerus (500 images). Results demonstrate significant benefits, with an average improvement of 14.9% in Fréchet Inception Distance, 9.7% in Improved Precision, and 2.3% in Improved Recall compared to state-of-the-art (SOTA) few-shot and data-limited natural image synthesis methods. Unlike other models, AnatoDiff consistently generates anatomically correct images with accurate shapes. Additionally, a ResNet-50 classifier trained on AnatoDiff-generated images shows a 2.1% to 5.3% increase in F1-score, compared to being trained on SOTA diffusion images, across 500 to 10,000 samples. A survey with 10 medical professionals reveals that images generated by AnatoDiff are challenging to distinguish from real ones, with a Matthews correlation coefficient of 0.277 and Fleiss' Kappa of 0.126, highlighting the effectiveness of AnatoDiff in generating high-quality, anatomically accurate radiographs. Our code is available at https://github.com/KawaiYung/AnatoDiff.