Feature-Based Pipeline for Improving Unsupervised Anomaly Segmentation on Medical Images

Abstract

Unsupervised methods for anomaly segmentation are promising for computer-aided diagnosis since they can increase the robustness of medical systems and do not require large annotated datasets. In this work, we propose a simple yet effective two-stage pipeline for improving the performance of existing anomaly segmentation methods. The first stage is used for better anomaly localization and false positive rate reduction. For this stage, we propose the PatchCore3D method, which is based on the PatchCore algorithm and a backbone, pre-trained on 3D medical images. Any existing anomaly segmentation method can be used at the second stage for the precise anomaly segmentation in the region suggested by PatchCore3D. We evaluate PatchCore3D and the proposed pipelines in combination with six top-performing anomaly segmentation methods of different types. We use brain MRI datasets, testing healthy subjects against subjects with brain tumors. Using PatchCore3D pipeline with every considered anomaly segmentation method increases segmentation AUROC almost twice by better anomaly localization.