DE-AE: A Dual-Encoder-Based Auto-encoder Framework with Improved Transformer for Anomaly Detection in Medical Imaging

Anomaly detection plays a crucial role in medical imaging because it does not require labeled samples. Existing anomaly detection methods are primarily based on image reconstruction techniques. However, image reconstruction-based methods face the issue of identity mapping. To solve the above issue, this paper presents a novel dual-encoder-based auto-encoder framework, which comprises two encoders sharing parameters and a decoder. The first encoder receives the raw medical image, while the second takes a synthesized noise image as an auxiliary input, aiding the first in better learning the manifold of normal samples. Further, an improved transformer is employed for modeling high-level features to mitigate the deficiencies of Convolutional Neural Networks (CNNs) in long-distance feature relationships. However, traditional Transformers bring a quadratic computational complexity. To balance model performance and computational cost, the framework uses CNNs to extract local features and the improved Transformer for modeling deep embedding features. The proposed method is validated on two public datasets, and the experimental results demonstrate the effectiveness of our approach.