Artifact Identification in Digital Pathology from Weak and Noisy Supervision with Deep Residual Networks

Computer-aided diagnosis in digital pathology often relies on the accurate quantification of different indicators using image analysis. However, tissue and slide processing can create various types of image artifacts: blur, tissue-fold, tears, ink stains, etc. On the basis of rough annotations, we develop a deep residual network method for artifact detection and segmentation in H&E and IHC slides, so that they can be removed from further image processing and quantification. Our results show that using detection (tile-based) or segmentation (pixel-based) networks (or a combination of both) can successfully find areas as large as possible of tissue with no artifact for further processing. We analyze how changes in the network architecture and in the data pre-processing influence the learning capability of the network. Networks were trained on the Hydra cluster of the ULB and VUB universities.