Fast Unsupervised Brain Anomaly Detection and Segmentation with Diffusion Models

Deep generative models have emerged as promising tools for detecting arbitrary anomalies in data, dispensing with the necessity manual labelling. Recently, autoregressive transformers achieved state-of-the-art performance anomaly detection medical imaging. Nonetheless, these still some intrinsic weaknesses, such requiring images to be modelled 1D sequences, accumulation of errors during sampling process, and significant inference times associated transformers. Denoising diffusion probabilistic are a class non-autoregressive recently shown produce excellent samples computer vision (surpassing Generative Adversarial Networks), achieve log-likelihoods that competitive while having relatively fast times. Diffusion can applied latent representations learnt by autoencoders, making them easily scalable great candidates application high dimensional images. Here, we propose method based on detect segment brain By training healthy data then exploring its reverse steps across Markov chain, identify anomalous areas space hence pixel space. Our compared approaches series experiments 2D CT MRI involving synthetic real pathological lesions much reduced times, their usage clinically viable.