MADGAN: unsupervised medical anomaly detection GAN using multiple adjacent brain MRI slice reconstruction

Area of Science:

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

Background:

• Unsupervised learning in medical imaging typically reconstructs single slices, limiting detection of diseases with subtle anatomical anomalies like Alzheimer's disease (AD).
• Existing methods struggle to assess disease stages, multiple disease types, or multi-sequence MRI scans simultaneously.
• There is a need for unsupervised anomaly detection methods that consider slice continuity and multi-modal data.

Purpose of the Study:

• To develop a novel unsupervised method for detecting various brain anomalies, including Alzheimer's disease and brain metastases, using multi-sequence MRI.
• To improve the detection of subtle anatomical anomalies and hyper-intense lesions by considering the continuity between adjacent MRI slices.
• To evaluate the method's performance across different disease stages and types, and on different MRI sequences.

Main Methods:

• Proposed unsupervised Medical Anomaly Detection Generative Adversarial Network (MADGAN), a two-step approach utilizing GAN-based reconstruction of adjacent brain MRI slices.
• Employed Wasserstein loss with Gradient Penalty and a secondary loss for training on healthy brain MRI slices to reconstruct subsequent slices.
• Utilized self-attention mechanism within MADGAN for enhanced anomaly detection on T1-weighted and contrast-enhanced T1 MRI scans.

Main Results:

• MADGAN successfully detected early-stage Alzheimer's disease (Mild Cognitive Impairment) with an AUC of 0.727 and late-stage AD with an AUC of 0.894 on T1-weighted MRI scans.
• The method achieved an AUC of 0.921 in detecting brain metastases on contrast-enhanced T1 MRI scans.
• Demonstrated reliable prediction of healthy brain slices and accurate discrimination of unseen abnormal scans, validating the reconstruction and diagnosis steps.

Conclusions:

• MADGAN, as the first unsupervised method for multi-sequence MRI analysis, reliably predicts healthy slices and detects subtle anatomical anomalies and lesions.
• The approach effectively diagnoses conditions like Alzheimer's disease and brain metastases, showcasing its potential for various disease detection.
• This method mimics physician diagnostic strategies by learning from extensive healthy data, offering a new paradigm for unsupervised anomaly detection in medical imaging.