Anomaly Detection in Pediatric and Adults Brain MRI with Generative Model

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology
• Machine Learning for Anomaly Detection

Background:

• Magnetic Resonance Imaging (MRI) is crucial for brain structure analysis but misses 5-10% of pathologies.
• Automating anomaly detection in MRI is necessary to reduce radiologist workload and improve diagnostic accuracy.

Purpose of the Study:

• To propose and evaluate a Denoising Diffusion Probabilistic Model (DDPM) for automated anomaly detection in brain MRI.
• To assess the efficacy of DDPM in differentiating healthy from unhealthy MRI scans and localizing abnormalities.

Main Methods:

• Utilized a Denoising Diffusion Probabilistic Model (DDPM) trained to reconstruct healthy brain MRI scans.
• Generated anomaly maps from input MRIs to highlight deviations from healthy patterns.
• Compared DDPM performance against other generative models, including Generative Adversarial Networks (GANs).

Main Results:

• The DDPM model correctly classified all test patients as either healthy or unhealthy.
• Generated anomaly maps effectively localized regions of interest indicative of pathologies.
• Achieved a Dice coefficient of 0.376, demonstrating superior performance compared to GANs.

Conclusions:

• DDPM shows significant promise as an effective tool for automated anomaly detection in brain MRI.
• The model's ability to generate anomaly maps aids in identifying subtle pathologies missed by conventional MRI analysis.
• DDPM offers an advancement over existing generative models for improving diagnostic efficiency and accuracy in neuroradiology.