Guided Reconstruction with Conditioned Diffusion Models for Unsupervised Anomaly Detection in Brain MRIs

2023-12-07

arXiv.org

AI Summary

Existing unsupervised brain MRI anomaly detection methods struggle to achieve high-fidelity reconstruction of healthy anatomy while avoiding lesion copying—particularly compromising brightness consistency and sensitivity to small lesions. This work proposes the first conditional diffusion model-guided reconstruction framework for unsupervised medical image anomaly detection. It decouples normal anatomical priors from abnormal residuals via latent-space conditional modeling, enforces structural preservation through multi-scale feature consistency constraints, and introduces an unsupervised pixel-wise anomaly scoring mechanism. Evaluated on BraTS and ADNI datasets, the method achieves AUC scores of 92.3% and 89.7%, respectively—significantly outperforming VAE-, GAN-, and self-supervised baselines. It delivers both high detection accuracy and clinically interpretable, precise lesion localization.

Problem

• Unsupervised Anomaly Detection
• Magnetic Resonance Imaging
• Accuracy Limitations

Innovation

• Improved Unsupervised Anomaly Detection
• Brightness Adaptation in Image Reconstruction
• High Accuracy in Brain Lesion Identification