BigpictureQC: multi-stain, multi-species artifact segmentation for whole-slide quality control
Abstract
Background: Artifacts in digital pathology whole-slide images (WSIs) cause diagnostic delay and quality issues, yet quality control (QC) in most pathology laboratories remains manual, subjective, and difficult to scale. We present BigpictureQC, a deep learning framework for WSI quality assessment based on per-class pixel-level segmentation of six visually distinct artifact and marking classes.
Methods: BigpictureQC was trained on 268 WSIs from five data sources and evaluated on >9,200 WSIs spanning human and preclinical samples, diverse tissue types, hematoxylin and eosin (H&E) and immunohistochemical (IHC) stains, and multiple scanner models. Patch-level segmentation was evaluated on 569 WSIs from seven independent cohorts using the Dice similarity coefficient for pixel-level segmentation; slide-level artifact detection was assessed on an independent dataset of 315 dermatopathology WSIs using F1 scores. A quantitative slide-level quality score was derived from artifact coverage and evaluated against manual quality ratings using Spearman rank correlation and AUC. Clinical impact was evaluated in a retrospective cohort of 3,686 diagnostic slides with technician rescan decisions and a prospective cohort of 4,426 unreviewed slides from routine diagnostics.
Results: BigpictureQC achieved a mean Dice score of 0.93 on the evaluation set, outperforming academic and commercial baselines across all artifact classes and domains (H&E, IHC, clinical, preclinical). Slide-level artifact detection yielded a micro-averaged F1 score of 0.93. The derived quality score strongly correlated with manual ratings and achieved an AUC of 0.978 for identifying poor-quality slides. In clinical workflow analyses, automated quality control identified 93% of poor-quality slides, compared with 35% detected by routine manual inspection. In the prospective-style cohort, 8.2% of slides were identified as having substantial quality issues that would otherwise have reached pathologists without prior review.
Conclusions: BigpictureQC provides the first publicly available automated QC framework validated across human and preclinical digital pathology, encompassing multiple tissue types, H&E and IHC stains, and diverse scanner models. It substantially outperforms existing manual and automated approaches in detecting poor-quality slides, and demonstrates strong potential to improve laboratory efficiency by enabling systematic pre-screening of all digitized slides at the point of scanning. The code, inference pipeline and trained model weights are publicly released.
Related
- Automated thigh muscle segmentation using foundation models: validation and impact of segmentation quality
- Task-Specific Quality Gating for Retinal Optical Coherence Tomography B-Scans: Learned Representations Over Scalar Metrics in Choroid Segmentation
- Species-dependent variation in flavor profile and sensory quality of green tea under aromatic woody intercropping systems.
- The Effect of Alpha-Lipoic Acid on Sperm Cryobiology and Post-Thaw Quality of Sperm in Livestock Species
- AI-driven quality enhancement of medical datasets: a closed-loop framework of synthesis, annotation, quality control, and integration
BibTeX
@article{damato2026bigpictureqc,
title = {{BigpictureQC: multi-stain, multi-species artifact segmentation for whole-slide quality control}},
author = {Marina D’Amato and Holger Hoefling and Erio Barale-Thomas and Freya Molenberghs and Pieter Van Nuffel and Anna Bodén and Paul van Diest and Nikolas Stathonikos and Sara Rombouts and Ivan R. Slootweg and Geert Litjens and Jeroen van der Laak and Francesco Ciompi},
journal = {BMC Medicine},
year = {2026},
month = oct,
doi = {10.1186/s12916-026-05257-1},
url = {https://link.springer.com/article/10.1186/s12916-026-05257-1},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - BigpictureQC: multi-stain, multi-species artifact segmentation for whole-slide quality control
AU - Marina D’Amato
AU - Holger Hoefling
AU - Erio Barale-Thomas
AU - Freya Molenberghs
AU - Pieter Van Nuffel
AU - Anna Bodén
AU - Paul van Diest
AU - Nikolas Stathonikos
AU - Sara Rombouts
AU - Ivan R. Slootweg
AU - Geert Litjens
AU - Jeroen van der Laak
AU - Francesco Ciompi
PY - 2026
DA - 2026/10/30
JO - BMC Medicine
DO - 10.1186/s12916-026-05257-1
UR - https://link.springer.com/article/10.1186/s12916-026-05257-1
AB - Background: Artifacts in digital pathology whole-slide images (WSIs) cause diagnostic delay and quality issues, yet quality control (QC) in most pathology laboratories remains manual, subjective, and difficult to scale. We present BigpictureQC, a deep learning framework for WSI quality assessment based on per-class pixel-level segmentation of six visually distinct artifact and marking classes. Methods: BigpictureQC was trained on 268 WSIs from five data sources and evaluated on >9,200 WSIs spanning human and preclinical samples, diverse tissue types, hematoxylin and eosin (H&E) and immunohistochemical (IHC) stains, and multiple scanner models. Patch-level segmentation was evaluated on 569 WSIs from seven independent cohorts using the Dice similarity coefficient for pixel-level segmentation; slide-level artifact detection was assessed on an independent dataset of 315 dermatopathology WSIs using F1 scores. A quantitative slide-level quality score was derived from artifact coverage and evaluated against manual quality ratings using Spearman rank correlation and AUC. Clinical impact was evaluated in a retrospective cohort of 3,686 diagnostic slides with technician rescan decisions and a prospective cohort of 4,426 unreviewed slides from routine diagnostics. Results: BigpictureQC achieved a mean Dice score of 0.93 on the evaluation set, outperforming academic and commercial baselines across all artifact classes and domains (H&E, IHC, clinical, preclinical). Slide-level artifact detection yielded a micro-averaged F1 score of 0.93. The derived quality score strongly correlated with manual ratings and achieved an AUC of 0.978 for identifying poor-quality slides. In clinical workflow analyses, automated quality control identified 93% of poor-quality slides, compared with 35% detected by routine manual inspection. In the prospective-style cohort, 8.2% of slides were identified as having substantial quality issues that would otherwise have reached pathologists without prior review. Conclusions: BigpictureQC provides the first publicly available automated QC framework validated across human and preclinical digital pathology, encompassing multiple tissue types, H&E and IHC stains, and diverse scanner models. It substantially outperforms existing manual and automated approaches in detecting poor-quality slides, and demonstrates strong potential to improve laboratory efficiency by enabling systematic pre-screening of all digitized slides at the point of scanning. The code, inference pipeline and trained model weights are publicly released.
SN - 1741-7015
ER -