UniTransAD: Unified Translation Framework for Anomaly Detection in Brain MRI

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Neuroscience

Background:

• Unsupervised anomaly detection (UAD) in brain MRI is vital for early disease diagnosis.
• Existing UAD methods struggle with generalization across different diseases, MRI sequences, and data missingness.
• Reconstruction-based methods often miss subtle anomalies, and translation methods lack input flexibility.

Purpose of the Study:

• To develop a unified and flexible framework for unsupervised anomaly detection in diverse brain MRI data.
• To address the limitations of current UAD methods in terms of generalization and detection of subtle anomalies.
• To introduce a robust and adaptable solution for clinical anomaly detection in varied healthcare settings.

Main Methods:

• Proposed UniTransAD, a unified translation-based anomaly detection framework.
• Introduced a unified cyclic-translation inference paradigm with content-style disentanglement for diverse MRI inputs.
• Developed a Dynamic Style Prototype Memory (DSPM) for flexible and robust cyclic inference.
• Implemented a dual-level detection mechanism combining pixel-level translation errors and feature-level dissimilarities.

Main Results:

• UniTransAD demonstrated superior flexibility and significantly outperformed state-of-the-art methods on the comprehensive Brain-OmniA dataset.
• The framework effectively handles diverse brain MRI inputs, including various diseases and sequences.
• The dual-level detection mechanism enhanced detection specificity for improved anomaly identification.

Conclusions:

• UniTransAD provides a robust, flexible, and generalizable solution for unsupervised anomaly detection in heterogeneous clinical environments.
• The proposed framework advances the capability for early and accurate diagnosis through improved brain MRI analysis.
• The developed Brain-OmniA dataset facilitates rigorous evaluation of UAD generalization beyond disease-specific datasets.