AnoDDPM: Anomaly Detection with Denoising Diffusion Probabilistic Models using Simplex Noise

Abstract

Generative models have been shown to provide a powerful mechanism for anomaly detection by learning to model \nhealthy or normal reference data which can subsequently \nbe used as a baseline for scoring anomalies. In this \nwork we consider denoising diffusion probabilistic models (DDPMs) for unsupervised anomaly detection. DDPMs \nhave superior mode coverage over generative adversarial \nnetworks (GANs) and higher sample quality than variational autoencoders (VAEs). However, this comes at the \nexpense of poor scalability and increased sampling times \ndue to the long Markov chain sequences required. We observe that within reconstruction-based anomaly detection \na full-length Markov chain diffusion is not required. This \nleads us to develop a novel partial diffusion anomaly detection strategy that scales to high-resolution imagery, named \nAnoDDPM. A secondary problem is that Gaussian diffusion fails to capture larger anomalies; therefore we develop \na multi-scale simplex noise diffusion process that gives control over the target anomaly size. AnoDDPM with simplex \nnoise is shown to significantly outperform both f-AnoGAN \nand Gaussian diffusion for the tumorous dataset of 22 T1- \nweighted MRI scans (CCBS Edinburgh) qualitatively and \nquantitatively (improvement of +25.5% Sørensen–Dice coefficient, +17.6% IoU and +7.4% AUC).