Equipping Computational Pathology Systems with Artifact Processing Pipelines: A Showcase for Computation and Performance Trade-offs

Histopathology is a gold standard for cancer diagnosis under a microscopicexamination. However, histological tissue processing procedures result inartifacts, which are ultimately transferred to the digitized version of glassslides, known as whole slide images (WSIs). Artifacts are diagnosticallyirrelevant areas and may result in wrong deep learning (DL) algorithmspredictions. Therefore, detecting and excluding artifacts in the computationalpathology (CPATH) system is essential for reliable automated diagnosis. In thispaper, we propose a mixture of experts (MoE) scheme for detecting five notableartifacts, including damaged tissue, blur, folded tissue, air bubbles, andhistologically irrelevant blood from WSIs. First, we train independent binaryDL models as experts to capture particular artifact morphology. Then, weensemble their predictions using a fusion mechanism. We apply probabilisticthresholding over the final probability distribution to improve the sensitivityof the MoE. We developed DL pipelines using two MoEs and two multiclass modelsof state-of-the-art deep convolutional neural networks (DCNNs) and visiontransformers (ViTs). DCNNs-based MoE and ViTs-based MoE schemes outperformedsimpler multiclass models and were tested on datasets from different hospitalsand cancer types, where MoE using DCNNs yielded the best results. The proposedMoE yields 86.15% F1 and 97.93% sensitivity scores on unseen data, retainingless computational cost for inference than MoE using ViTs. This bestperformance of MoEs comes with relatively higher computational trade-offs thanmulticlass models. The proposed artifact detection pipeline will not onlyensure reliable CPATH predictions but may also provide quality control.