Anomaly Detection in Medical Images Using Encoder-Attention-2Decoders Reconstruction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Anomaly detection (AD) in medical applications offers a cost-effective alternative to manual data labeling.
• Feature reconstruction-based AD methods face challenges including domain gap issues and underutilized decoder potential.

Purpose of the Study:

• To introduce a novel method, Encoder-Attention-2Decoder (EA2D), for effective anomaly detection in medical imaging.
• To address the domain gap in pre-trained encoders and enhance decoder exploration for improved AD performance.

Main Methods:

• EA2D employs a primary feature reconstruction task and an auxiliary transformation-consistency contrastive learning task.
• A self-attention skip connection augments reconstruction quality for normal cases.
• Dual decoders are utilized to reconstruct dual image views, mitigating over-reconstruction of anomalies.

Main Results:

• EA2D demonstrates superior performance across four medical image modalities.
• The method effectively reduces the domain gap between natural and medical images.
• Enhanced reconstruction quality of normal cases improves anomaly differentiation.

Conclusions:

• EA2D offers a significant advancement in medical anomaly detection.
• The proposed method effectively overcomes limitations of existing feature reconstruction techniques.
• EA2D shows promise for various medical imaging applications, with code availability for further research.