Denoising Autoencoders for Unsupervised Anomaly Detection in Brain MRI
Antanas Kascenas, Nicolas Pugeault, Alison Q O'Neil
Pathological brain lesions exhibit diverse appearance in brain images, making it difficult to design specialized detection solutions due to the lack of comprehensive data and annotations. Thus, in this work we tackle unsupervised anomaly detection, using only healthy data for training with the aim of detecting unseen anomalies at test time. Many current approaches employ autoencoders with restrictive architectures (i.e. containing information bottlenecks) that tend to give poor reconstructions of not only the anomalous but also the normal parts of the brain. Instead, we investigate classical denoising autoencoder models that do not require bottlenecks and can employ skip connections to give high resolution fidelity. We design a simple noise generation method of upscaling low-resolution noise that enables high-quality reconstructions, reducing false positive noise in reconstruction errors. We find that with appropriate noise generation, denoising autoencoder reconstruction errors generalize to hyperintense lesion segmentation and can reach state of the art performance for unsupervised tumor detection in brain MRI data, beating more complex methods such as variational autoencoders. We believe this provides a strong and easy-to-implement baseline for further research into unsupervised anomaly detection.
Thursday 7th July
Oral Session 2.2 - 13:20 - 14:00 (UTC+2)