U²AD: Uncertainty-based unsupervised anomaly detection framework for detecting T2 hyperintensity in MRI spinal cord.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Spinal cord T2 hyperintensities on MRI are key biomarkers for conditions like degenerative cervical myelopathy (DCM).
• Current clinical diagnosis relies on manual evaluation, which is time-consuming and subjective.
• Supervised deep learning methods require extensive annotated datasets, limiting their clinical applicability.

Purpose of the Study:

• To develop an unsupervised anomaly detection (UAD) framework, U²AD, to identify spinal cord T2 hyperintensities.
• To overcome limitations of existing UAD methods, including domain shifts and task conflict.
• To eliminate the need for abnormal data annotations in lesion detection.

Main Methods:

• Proposed an Uncertainty-based Unsupervised Anomaly Detection (U²AD) framework using a Vision Transformer.
• Implemented a "mask-and-reconstruction" paradigm trained and tested on the same clinical dataset.
• Introduced an uncertainty-guided masking strategy employing Monte-Carlo inference for epistemic and aleatoric uncertainty estimation.

Main Results:

• U²AD demonstrated superior performance in both patient-level identification and segment-level localization of spinal cord T2 hyperintensities compared to existing UAD methods.
• The uncertainty-guided strategy effectively balanced normal reconstruction and anomaly detection.
• Improved normal representation learning while enhancing sensitivity to anomalies.

Conclusions:

• U²AD establishes a new benchmark for uncertainty-guided UAD in medical imaging.
• The framework offers a promising alternative for automated detection of spinal cord lesions, aiding in DCM diagnosis.
• This approach reduces reliance on manual annotation and supervised learning, facilitating broader clinical adoption.