Unsupervised anomaly detection in brain MRI via disentangled anatomy learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Brain MRI lesion detection is crucial but challenging due to lesion diversity and imaging variability.
• Current unsupervised methods struggle with generalizability across different MRI modalities and centers.
• Existing models are limited by abnormal residuals in reconstructed pseudo-healthy images (PHIs).

Purpose of the Study:

• To develop a novel unsupervised framework for robust brain MRI lesion detection.
• To enhance the generalizability and performance of anomaly detection in brain MRIs.
• To address limitations of current methods, including restricted generalizability and performance constraints.

Main Methods:

• Proposed a new PHI reconstruction framework with two novel modules: disentangled representation and edge-to-image restoration.
• The disentangled representation module separates imaging information from anatomical images for improved generalizability.
• The edge-to-image restoration module reconstructs high-quality PHIs by restoring anatomical details from edge information.

Main Results:

• The proposed method achieved absolute improvements of +18.32% in average precision and +13.64% in Dice similarity coefficient.
• Outperformed 17 state-of-the-art methods across nine public datasets (4,443 patients' MRIs).
• Demonstrated superior performance in multi-modality and multi-center brain MRI lesion detection.

Conclusions:

• The novel framework significantly improves unsupervised brain MRI lesion detection.
• The method offers enhanced generalizability and performance compared to existing techniques.
• This approach holds promise for more accurate and reliable clinical diagnosis of brain anomalies.