BigpictureQC detects artifacts in whole-slide pathology images across cohorts

BigpictureQC is a deep-learning framework that segments six classes of artifacts and markings in whole-slide images. It was trained on 268 slides and evaluated across more than 9,200 human and preclinical slides, including H&E and immunohistochemical stains and multiple scanner models. Segmentation performance was assessed on 569 slides from seven independent cohorts; slide-level detection was tested on 315 dermatopathology slides. Clinical workflow analyses included 3,686 diagnostic slides retrospectively and 4,426 unreviewed routine slides prospectively.

The model achieved a mean Dice score of 0.93 for pixel-level segmentation and a micro-averaged F1 score of 0.93 for slide-level artifact detection. Its quality score had an AUC of 0.978 for identifying poor-quality slides. In the retrospective workflow analysis, automated QC identified 93% of poor-quality slides, compared with 35% detected by routine manual inspection. In the prospective-style cohort, 8.2% of slides had substantial quality issues that would otherwise have reached pathologists without prior review.

The findings support further evaluation of automated prescreening as an aid to digital pathology quality control. The study assessed image quality and workflow detection, not diagnostic accuracy or patient outcomes; implementation would require validation in local workflows. Code, an inference pipeline, and model weights were publicly released. The project received funding through the BigPicture consortium, including EU and EFPIA support.