A Unified Conditional Framework for Unsupervised Anomaly Detection

Unsupervised anomaly detection methods based on Generative Adversarial Networks (GANs) have gained momentum tremendous in medical applications. However, due to the diverse range of anomalies encountered in real-world clinical settings, the utilization of GANs remains constrained to addressing one specific pathology per model. This paper introduces an innovative unsupervisedapproach to pulmonary anomaly detection through a unified conditional frame-work. The model consists of two Residual Attention Conditional networks: a GAN and an Encoder. Tailored to the input condition, such as the specified dataset or region of interest (ROI), the GAN generates realistic samples in the target style, while the encoder ensures fast mapping of normal images to their latent representations. Notably, our Residual Attention Conditional GAN (RA-CGAN) framework is exclusively trained on normal data in a fully unsupervisedmanner. During inference, our model adeptly detects anomalous images acrossvarious contexts, eliminating the need for distinct models for each dataset or ROI.This unified approach not only streamlines training processes but also harnesses joint information, significantly enhancing generalization capabilities. Additionally, the model integrates Depthwise Separable Convolution (DSC) to reduce computational resources and leverages the Convolutional Block Attention Module (CBAM) to focus on relevant image regions without increasing model complexity. The efficiency of our framework is demonstrated on pneumonia and lung cancer datasets. Benchmarking results on MVTec-AD datasets illustrate its state-of-the-art performance at the image level, contributing to efficient unsupervised anomaly detection.