A Patch-based Student-Teacher Pyramid Matching Approach to Anomaly Detection in 3D Magnetic Resonance Imaging

Johannes Schwarz, Lena Will, Jörg Wellmer, Axel Mosig

Keywords: magnetic resonance imaging, anomaly detection, semi-supervised learning, student-teacher feature pyramid matching, voxel and image-level detection

Abstract: Anomaly detection on 3D magnet resonance images (MRI) is of high medical relevance in the context of detecting lesions associated with different diseases. Yet, reliable anomaly detection in MRI images involves major challenges, specifically taking into account information in 3D, and the need to localize relatively small and subtle abnormalities within the context of whole organ MRIs. In this paper, a top-down approach, which uses student-teacher feature pyramid matching (STFPM) for detecting anomalies at image and voxel level, is applied to 3D brain MRI inputs. The combination of a 3D patch based self-supervised pre-training and axial-coronal-sagittal (ACS) convolutions pushes the performance above that of f-AnoGAN (bottom-up). The evaluation is based on a tumor dataset.

Submission Number: 205