Diffusion Models for Medical Anomaly Detection

AbstractIn medical applications, weakly supervised anomaly detection methods are of great interest, as only image-level annotations required for training. Current mainly rely on generative adversarial networks or autoencoder models. Those models often complicated to train have difficulties preserve fine details in the image. We present a novel method based denoising diffusion implicit combine deterministic iterative noising and scheme with classifier guidance image-to-image translation between diseased healthy subjects. Our generates very detailed maps without need complex training procedure. evaluate our BRATS2020 dataset brain tumor CheXpert detecting pleural effusions.KeywordsAnomaly detectionDiffusion modelsWeak supervision