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

TL;DR

This work tackles the detrimental impact of histological artifacts on CPATH by introducing a Mixture of Experts (MoE) approach and multiclass baselines to detect five artifacts in WSIs. The MoE framework trains five binary experts (one per artifact) and fuses their outputs with a meta-threshold $t_s$ to maximize artifact-free sensitivity, while also providing artifact segmentation and QC outputs. Evaluations on internal EMC_dev data and external OoD cohorts show that DCNN-based MoE offers the best overall performance, albeit with higher computational cost, whereas multiclass models provide faster inference with a modest performance trade-off. The proposed artifact-processing pipeline yields segmentation maps, artifact reports, artifact-free ROI masks, and artifact-free WSIs, enabling more reliable CPATH predictions and built-in quality control across diverse tissue types and scanners.

Abstract

Histopathology is a gold standard for cancer diagnosis under a microscopic examination. However, histological tissue processing procedures result in artifacts, which are ultimately transferred to the digitized version of glass slides, known as whole slide images (WSIs). Artifacts are diagnostically irrelevant areas and may result in wrong deep learning (DL) algorithms predictions. Therefore, detecting and excluding artifacts in the computational pathology (CPATH) system is essential for reliable automated diagnosis. In this paper, we propose a mixture of experts (MoE) scheme for detecting five notable artifacts, including damaged tissue, blur, folded tissue, air bubbles, and histologically irrelevant blood from WSIs. First, we train independent binary DL models as experts to capture particular artifact morphology. Then, we ensemble their predictions using a fusion mechanism. We apply probabilistic thresholding over the final probability distribution to improve the sensitivity of the MoE. We developed DL pipelines using two MoEs and two multiclass models of state-of-the-art deep convolutional neural networks (DCNNs) and vision transformers (ViTs). DCNNs-based MoE and ViTs-based MoE schemes outperformed simpler multiclass models and were tested on datasets from different hospitals and cancer types, where MoE using DCNNs yielded the best results. The proposed MoE yields 86.15% F1 and 97.93% sensitivity scores on unseen data, retaining less computational cost for inference than MoE using ViTs. This best performance of MoEs comes with relatively higher computational trade-offs than multiclass models. The proposed artifact detection pipeline will not only ensure reliable CPATH predictions but may also provide quality control.

Figures (13)

• Figure 1: An overview of computational pathology (CPATH) system equipped with artifact processing pipeline. Whole slide images (WSIs) are split into small sub-images (patches) to make them computationally tractable for deep learning (DL) models. These patches are fed to a mixture of experts (MoE) or multiclass models composed of state-of-the-art DL architectures to perform different CPATH classification tasks. Only patches with histological relevance can flow further for the downstream tasks. Finally, predictions are post-processed to produce different outcomes, such as a segmentation map, artifact report for quality control, region of interest mask, and artifact-free WSI for the diagnostic or prognostic algorithm to make a final clinical prediction.
• Figure 2: Examples of artifact-free and artifact-classes patches in our prepared patch-based dataset $\mathcal{D}$ from EMC$_{dev}$, and extracted at 40x magnification.
• Figure 3: An overview of the mixture of experts (MoE) formation for artifact detection. Five base learners (either MobileNetv3 or ViT-Tiny deep learning architectures) are trained on overlapping sub-datasets to learn the distinct morphology of each artifact. Labels are transformed to take the artifact class as a negative class. A fusion function integrates output from all experts to form a predictive probability distribution for the final prediction. A meta-learned probability threshold is applied to maximize the sensitivity of the MoE.
• Figure 4: Overview of deep learning pipeline emphasizing the post-processing stage during the inference.Pre-processing: The whole slide image (WSI) is split, and every patch is stored with its corresponding coordinate. Inference: Every patch is assigned a label using a mixture of experts or multiclass DL models. Post-processing: The matrix-based filling method assigns a color to every pixel (in the down-scaled version of WSI) at the corresponding coordinate location. Post-processing provides: 1) Segmentation map; 2) Artifact report for quality control; 3) Artifact-free region of interest map, and 4) Artifact-refined WSI for computational analysis.
• Figure 5: ROC curves for deep learning pipelines over the validation subset. All plots highlight the area under the curves (AUC) score and best probability thresholds for maximizing F1 and sensitivity metrics.