Stable Diffusion Model-based Scintigraphy Image Synthesis: Data Augmentation Toward Enhanced Multiclass Thyroid Diagnosis

The objective of this study is to assess the efficacy of advanced augmentation techniques, such as stable diffusion, in improving the performance of deep learning models in the classification of scintigraphic thyroid images. In this retrospective study, 2983 anterior view scintigraphic images were collected and subsequently categorized into four thyroid conditions. Both stable diffusion and conventional augmentation techniques were utilized. The generated images, alongside real images, were used to train a ResNet101V2 architecture under six different training strategies. The strategies were assessed against external datasets to evaluate model performance in terms of accuracy, precision, recall, and F1-score. The use of synthetic data in training led to consistently superior performances compared to training with only real data. Specifically, the models trained with synthetic data augmentation demonstrated higher precision and recall. The incorporation of synthetic images generated via stable diffusion significantly enhanced the diagnostic capabilities of AI models in thyroid scintigraphy interpretation. This approach not only improves the classification accuracy but also provides a viable solution to the challenge of data scarcity in medical imaging.