IterMask3D: Unsupervised anomaly detection and segmentation with test-time iterative mask refinement in 3D brain MRI

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Unsupervised anomaly detection methods learn normal patterns to identify deviations as anomalies.
• Current methods corrupt images for training, risking information loss and false positives.
• 3D brain MRI analysis requires robust methods for anomaly detection and segmentation.

Purpose of the Study:

• To introduce IterMask3D, an iterative spatial mask-refining strategy for unsupervised anomaly detection and segmentation in 3D brain MRI.
• To reduce false positives in anomaly detection by minimizing information loss during the learning process.
• To enhance reconstruction performance by incorporating high-frequency image content.

Main Methods:

• Iterative spatial masking and reconstruction of 3D brain MRI data.
• Mask shrinking based on reconstruction error to progressively unmask normal regions.
• Utilizing high-frequency image content to guide reconstruction of masked areas.

Main Results:

• IterMask3D effectively detects synthetic and real-world imaging artifacts in 3D brain MRIs.
• The method demonstrates strong performance in segmenting various pathological lesions across multiple MRI sequences.
• Consistent effectiveness shown in extensive experiments, reducing false positives compared to prevailing methods.

Conclusions:

• IterMask3D offers a significant advancement in unsupervised anomaly detection and segmentation for 3D brain MRI.
• The iterative mask-refining strategy enhances accuracy and reduces false positives.
• The method shows broad applicability for artifact detection and lesion segmentation in neuroimaging.