Zero-Shot Anomaly Detection in Medical Images Using Conditional Denoising Diffusion Models and Data Augmentation

Zero-shot anomaly detection holds significant importance in medical image analysis, particularly when newly emerging anomaly categories are absent from the training set. Traditional methods often perform poorly due to their inability to effectively classify and locate anomalous regions. In this study, we propose a method that combines the Denoising Diffusion Anomaly Detection (DDAD) model with data augmentation strategies for zero-shot anomaly detection in medical images. Prior to data training, we employ various data augmentation techniques, such as contrast enhancement, sharpening, rotation, and translation, to increase data diversity and enhance the model's ability to learn image features. This improves the model's generalization capability and robustness. Experimental validation on medical image datasets demonstrates that our method exhibits superior performance in zero-shot anomaly detection tasks, significantly outperforming the current state-of-the-art models. Our experiments on the BMAD benchmark for medical anomaly detection demonstrate significant improvements in zero-shot anomaly detection for the Brain, Liver, and RESC datasets. Specifically, the AUROC for anomaly classification increased by 1.67%, 16.56%, and 13.09%, respectively, while the AUROC for anomaly localization improved by 4.13%, 0.65%, and 5.85%, respectively,