Tissue-Preserving Artifact Quality Control for Whole-Slide Computational Pathology

Reliable whole-slide imaging (WSI) analysis depends on diagnostic models, but it also depends on image-quality control that can be assessed before those models are applied. High-sensitivity artifact masks may remove relevant tissue, especially in densely nucleated regions where benign histology resembles dust, folds, or preparation debris. GuardQC is a conservative preprocessing pipeline for artifact control and tissue preservation in computational pathology. The method uses a coarse-to-fine strategy: a high-recall proposal network identifies suspicious regions, and a dual-stream ResNet-18 verification model combines $64\times 64$ local texture with $512\times 512$ tissue context after dynamic multi-resolution alignment to 1.5  $\mu$ m/pixel. Iterative hard negative mining is used to improve specificity under noisy quality-control labels. Localized Otsu re-segmentation then contracts confirmed artifact regions, and deterministic PDE-based inpainting neutralizes artifacts without synthesizing new tissue content. On an expert-curated subset, GuardQC achieved 98.99% accuracy and 100% specificity (0/101 false positives; 95% CI, 96.4%-100.0%). In the evaluated scanner-stratified WSI cohort, the masked tissue percentage was 0.32%, compared with 0.94% for a deep learning baseline. Feature ablations showed recovery of representation consistency by more than 87% for CNN and contrastive models and by 64.10% for a self-supervised foundation model. These results support GuardQC as a quality-control layer for evaluated WSI preprocessing settings that require artifact control, scanner-aware analysis, tissue preservation, and inspectable processing.