ESC-DRKD: Enhanced skip connection-based direct reverse knowledge distillation for medical image anomaly detection

Image anomaly detection has emerged as a prevalent research area in medical diagnosis, focusing on using models trained exclusively on normal samples to identify and locate anomalous images during testing. While many image anomaly detection (AD) models exhibit remarkable performance on industrial datasets, they frequently struggle to effectively detect anomalies in medical datasets characterized by complex distributions. This is because most methods are prone to overgeneralization, resulting in ineffective restoration of normal image features. To tackle this challenge, we introduce a novel approach called Enhanced Skip Connection-Based Direct Reverse Knowledge Distillation (ESC-DRKD), specifically designed to facilitate anomaly detection and localization in medical images. ESC-DRKD consists of pre-trained teacher encoders, trainable projection layers, and student decoders. Firstly, pseudo abnormal images are generated using the CutPaste method. By extracting multi-scale features from both normal and pseudo abnormal images through the pretrained teacher encoders, projection layers are employed to project the features of pseudo abnormal images onto important information of normal features. These projected features are then added to the outputs of corresponding student decoders at each level through skip connections to enhance the restoration of normal features. Furthermore, the output of the final layer of the teacher encoders serves as the input to the student decoders, thereby preventing the loss of normal information. Extensive experiments on various public medical datasets are tested the effectiveness of our proposed method. The results demonstrate that ESC-DRKD outperforms the state-of-the-art AD models on the five medical datasets, achieving an average improvement of over 3.0 % AUROC for anomaly detection. Code is available at https://github.com/GE-123-cpu/ESC_DRKD