Unsupervised Artifact Detection for Whole Slide Images of Prostate Biopsies

High-quality image digitisation of histological slides is essential for digital pathology to facilitate diagnosis and to develop reliable computer-aided assistance systems. Currently, image quality control (QC) to identify artefacts that result from slide preparation, staining, or scanning is mainly conducted manually, which is time-consuming and prone to errors. The few existing automated approaches are limited to specific artefacts, require manual parameter tuning, or depend on large amounts of annotated data for training. Here, we propose an unsupervised and training-free method for artefact detection of digitised histological slides based on one-class classification. By assuming that deep representations from normal (in-distribution) tissue samples are dense and compact in feature space, artefacts can be identified by their distance to the mean in-distribution representation. The use of representations from convolutional neural networks pre-trained on natural images (ImageNet) avoids the need for the collection of annotated data and the costs of training. We demonstrate the efficacy of our approach for the detection of various types of artefacts (out-of-focus, pen marks, tissue folds, etc.) in an open dataset of prostate biopsy whole-slide images. We compare the proposed distance-based artefact detection with conventional task-specific artefact-detection methods as well as with approaches based on uncertainty estimation. Our method achieved a mean AUROC value of 0.82 on our test dataset consisting of 78 prostate biopsy whole-slide images annotated with five varieties of artefacts. Our proposed method can be utilised for slide-level QC as well as the identification of artefacts within a slide. It is a simple, efficient, yet robust method for automated QC of digitized histological slides.