High-Fidelity Artificial Medical Image Generation designed for Rare Tissues Using Diffusion Models Enhancing Diagnostic Accuracy and AI Training

Even common examples of medical pathologies that are not properly represented in medical imaging databases, like rare sarcomas and gliomas, represent a huge challenge to developing robust diagnostic models by AI methods. Generation Current generative models, such as GANs and VAEs, have shown many limitations in their performance, namely high rates of mode collapse, training instability, and failure to generate anatomically accurate rarely seen tissue structures. To overcome such a challenge, this paper proposes a conditional diffusion-based generative model aimed at imitating high-fidelity, clinically realistic imaging of rare tumors. The model also incorporates semantic conditioning with tumor masks and implementation of class embedding and adds the adversarial validation filtering to guarantee anatomic plausibility. The produced synthetic images were verified on The Cancer Imaging Archive (TCIA) Rare Tumor Subset, improving quantitatively on baseline models: FID was - 14.7, SSIM increased by 11.3 and PSNR rose by 2.9 dB when compared to GAN- and VAE-based models. In addition, using augmented datasets (real + synthetic) to retrain a ResNet-50 classifier enhanced classification accuracy and F1-score by 8 percent and 9 percent configurations, respectively, compared to classification based solely on real data. Such results emphasize the prospect of diffusion-based frameworks in alleviating the shortage of labeled data, improving the performance of rare-case diagnoses, and offering a scalable approach to medical AI training.