Unsupervised anomaly detection in medical imaging using aggregated normative diffusion

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Neuroscience Research

Background:

• Early detection of anomalies in brain magnetic resonance imaging (MRI) is crucial for medical diagnosis and treatment.
• Supervised machine learning for anomaly detection is limited by the need for extensive labeled data for specific pathologies.
• Unsupervised anomaly detection (UAD) offers a promising approach to identify a wider range of abnormalities by detecting deviations from normal patterns.

Purpose of the Study:

• To address the limitations of existing UAD methods in generalizing to diverse anomalies in multi-modal brain MRI data.
• To introduce and evaluate a novel UAD method named Aggregated Normative Diffusion (ANDi).

Main Methods:

• ANDi aggregates differences between predicted denoising steps and ground truth backwards transitions within Denoising Diffusion Probabilistic Models (DDPMs).
• The DDPMs used in ANDi are trained on pyramidal Gaussian noise.
• The proposed method was validated against four recent UAD baselines across three diverse brain MRI datasets.

Main Results:

• ANDi demonstrates substantial improvements over existing UAD baselines in detecting anomalies in brain MRI.
• The method exhibits increased robustness to various types of anomalies.
• Specifically, ANDi achieved up to a 44% improvement in Area Under the Precision-Recall Curve (AUPRC) for detecting multiple sclerosis (MS) lesions.

Conclusions:

• Aggregated Normative Diffusion (ANDi) represents a significant advancement in unsupervised anomaly detection for multi-modal brain MRI.
• The developed method offers enhanced accuracy and robustness, particularly beneficial for identifying neurological conditions like MS.
• ANDi has the potential to improve the early detection and diagnosis of a broader range of brain abnormalities.