Unsupervised Anomaly Detection for Medical Images Based on Multi-Hierarchical Feature Reconstruction

Unsupervised anomaly detection based on reconstruction is receiving focused research due to its low annotation requirement with gradually improving accuracy. However, the reconstruction quality and detection effect of existing methods still need to be improved. In this paper, we propose an unsupervised anomaly detection method based on multi-hierarchical feature reconstruction, specifically, we set the reconstruction object as the intermediate features of the network instead of the general image. First, we adopt Transformer as the backbone network and use an encoder pre-trained on a large-scale dataset to extract meaningful features. Second, we construct a decoder corresponding to the encoder, and the training goal is to make the intermediate features of the decoder approximate the intermediate features with those of the encoder. During training, to prevent overgeneralization, we propose a perturbation module to simulate anomalies to interfere with the decoder’s input. Experiments on the Hyper-Kvasir and BraTS 2021 dataset show that the proposed method can effectively improve the anomaly detection metrics, achieving the best performance with excellent stability.