Self-supervised Diffusion Model for Anomaly Segmentation in Medical Imaging

Abstract

A powerful mechanism for detecting anomalies in a self-supervised manner was demonstrated by model training on normal data, which can then be used as a baseline for scoring anomalies. Recent studies on diffusion models (DMs) have shown superiority over generative adversarial networks (GANs) and have achieved better-quality sampling over variational autoencoders (VAEs). Owing to the inherent complexity of the systems being modeled and the increased sampling times in the long sequence, DMs do not scale well to high-resolution imagery or a large amount of training data. Furthermore, in anomaly detection, DMs based on the Gaussian process do not control the target anomaly size and fail to repair the anomaly image, which led us to the development of a simplex diffusion and selective denoising ((SD) 2) model. (SD) 2 does not require a full sequence of Markov chains in image reconstruction for anomaly detection, which reduces the time complexity, samples from the simplex noise diffusion process that have control over the anomaly size and are trained to reconstruct the selective features that help to repair the anomaly. (SD) 2 significantly outperformed the publicly available Brats2021 and Phenomena detection from X-ray image datasets compared to the self-supervised model. The source code is made publicly available at https://github.com/MAXNORM8650/SDSquare.