henhaobio
Sign in

Scientific ReportsBioinformaticsPeer-reviewed journal article

QuPath: Open source software for digital pathology image analysis

Peter BankheadiD, Maurice Bernard LoughreyiD, José A. FernándeziD, Yvonne DombrowskiiD, Darragh G. McArtiD, Philip D. DunneiD, Stephen McQuaidiD, Ronan T. GrayiD, Liam MurrayiD, Helen G. ColemaniD, Jacqueline A. JamesiD, Manuel Salto‐TelleziD, Peter W. Hamilton

PDFDOIPreprintCode
In-depth full-text digest available · Jump to digest ↓

Abstract

QuPath is new bioimage analysis software designed to meet the growing need for a user-friendly, extensible, open-source solution for digital pathology and whole slide image analysis. In addition to offering a comprehensive panel of tumor identification and high-throughput biomarker evaluation tools, QuPath provides researchers with powerful batch-processing and scripting functionality, and an extensible platform with which to develop and share new algorithms to analyze complex tissue images. Furthermore, QuPaths flexible design make it suitable for a wide range of additional image analysis applications across biomedical research.

In-depth summary

Claude Opus 5.5 · max effort · 8 Nov 2026 · from the full text (6 Mar 2017) · CC BY 4.0

Open-source QuPath combines tile-based whole slide viewing, object hierarchies and trainable classifiers; it automated CD3, CD8, p53, PD-L1 and stromal scoring that stratified colon cancer survival.

Open-source bioimage tools such as ImageJ, Fiji, Icy and CellProfiler do not address the viewing and computing demands of gigapixel whole slide images, and existing open pathology tools only read slide formats, analyze cropped tiles or manage data, so users lacking commercial software downsampled images or relied on variable manual scoring. QuPath offers an extensible, scriptable desktop framework for whole-slide viewing, annotation, cell-level analysis and algorithm sharing.

Key results

  1. ShownScripted batch counting of DAB-positive cells per mm^2 in CD3 and CD8 tissue microarrays (TMAs) ran in minutes on a desktop workstation; median-split densities of both markers separated disease-specific survival, reproducing the known prognostic role of adaptive immunity in colorectal cancer.

    Under 4 minutes per biomarker for 21 whole slide images (>2000 cores per biomarker), counting about 1.2 million CD3+ and 0.6 million CD8+ cells on a Mac Pro (3.5 GHz 6-core Xeon E5, 32-GB RAM), with markup images exported for checking; log-rank p = 0.006 (CD3) and p = 0.007 (CD8) for disease-specific survival. · Results and Discussion; Software; Analysis of CD3 and CD8 IHC

  2. SuggestedAn interactively trained random trees classifier restricted p53 H-scoring to tumor epithelial cells; patients with aberrant (absent or diffuse intense) staining had poorer unadjusted disease-specific survival than those with intermediate, wild-type-like expression.

    Cell detection gives up to 33 measurements per cell; 16 were kept and supplemented with local cell density and Gaussian-smoothed neighbor features. H-score cutoffs of 10 and 160, picked by a pathologist after viewing cores ranked by score, defined aberrant groups; p = 0.003 vs intermediate (log-rank, unadjusted). · Results and Discussion; Statistical analysis; Analysis of p53 IHC

  3. ShownPD-L1 positivity was called from maximum DAB optical density within nucleus-expanded cell areas after three-way cell classification; higher percentages of positive cells were associated with better unadjusted disease-specific survival, with a dose-response across tertiles.

    Three-way random trees (epithelial, non-epithelial, other); score = percent positive cells per core with 'other' removed, patient score = median of up to three cores. Median cutoff of 1.46% positive cells gave p = 0.004 (log-rank, unadjusted), higher expression favorable; tertiles in Supplementary Materials. · Results and Discussion; Statistical analysis; Analysis of PD-L1 IHC

  4. ShownOn whole-face hematoxylin and eosin (H&E) sections, pathologist-trained superpixel texture classification estimated tumor stromal percentage, and a high stromal fraction marked lower unadjusted disease-specific survival in stage II patients, consistent with earlier visual-estimate studies.

    312 stage II slides (231 Aperio ScanScope, 81 Hamamatsu Nanozoomer); scripted per-image RGB background estimation; simple linear iterative clustering (SLIC) superpixels with hue, Haralick and smoothed neighbor features; random trees trained on 40 images (epithelium, stroma, other); median cutoff, log-rank p = 0.013 (unadjusted). · Results and Discussion; Image data acquisition; Analysis of tumor stromal percentage in H&E…

  5. SuggestedSplitting PD-L1 scores by classifier-assigned compartment indicated that PD-L1-positive cells in these colon tumors lie mainly outside the tumor epithelium.

    Reported qualitatively from an epithelial versus non-epithelial breakdown in Supplementary Fig. 4; no compartment-specific percentages or test statistics are given in the main text. · Results and Discussion; Analysis of PD-L1 IHC

Methods, data and reproducibility

QuPath is a cross-platform Java/JavaFX desktop application combining a multithreaded tile-based slide viewer with a hierarchical object model; scripted TMA dearraying, color deconvolution, cell detection, interactive random trees classification and superpixel texture analysis generated biomarker scores tested against colon cancer survival.

  • Population-based stage II/III colon adenocarcinoma cohort (89% of resected cases in two Northern Ireland Healthcare Trusts, 2004–2008) with registry-linked survival follow-up · human · n = 660 patients (392 stage II, 268 stage III)
  • TMAs of triplicate 1 mm tumor-center cores immunostained for CD3 (clone 2GV6) and CD8 (clone c8/144B), scanned on an Aperio ScanScope CS at 40X · human · n = 21 slide images and >2000 cores per marker
  • Second TMA set from the same cohort immunostained for p53 (clone DO-7), H-scored within classifier-identified tumor epithelial cells · human
  • TMAs from the same cohort immunostained for PD-L1 (clone SP142), scored as percentage of positive cells after three-way cell classification · human
  • Whole-face representative tumor H&E sections from stage II patients, scanned on Aperio ScanScope (231) or Hamamatsu Nanozoomer (81) · human · n = 312 patients/slides; 40 used for classifier training
Methods & tools
QuPath (Java 8, JavaFX interface), OpenSlide (whole slide image reading), ImageJ (as image-processing library), OpenCV random trees classifier and fast cell counting, QuPath TMA dearrayer, Color deconvolution with QuPath Estimate stain vectors, QuPath Simple tissue detection and Fast cell counts, QuPath Cell detection, Add intensity features and Add smoothed features, SLIC superpixel segmentation with Haralick texture features, Kaplan-Meier and log-rank tests (QuPath TMA data viewer), R 3.2.2 with Survival package 2.38–3, Immunohistochemistry clones: CD3 2GV6, CD8 c8/144B, p53 DO-7, PD-L1 SP142
Reagents & environment
Java 8/JavaFX; OpenSlide, ImageJ, OpenCV; R 3.2.2, survival 2.38–3. Antibody clones: CD3 2GV6 (Ventana BenchMark, CC1 32 min), CD8 c8/144B (Dako, Leica Bond III, ER2 20 min, 1/50), p53 DO-7 (Dako, ER2 30 min, 1/100), PD-L1 SP142 (Ventana, CC1 24 min).
Assay conditions & compute
Mac Pro (3.5 GHz 6-core Intel Xeon E5, 32-GB RAM); under 4 minutes per biomarker for 21 TMA slides. TMAs scanned on Aperio ScanScope CS at 40X (0.25 μm/pixel); H&E slides on Aperio (231) or Hamamatsu Nanozoomer (81), 0.231–0.253 μm/pixel.
Reproducibility gaps
No data-availability statement: slide images, patient-level scores and clinical data, batch scripts and trained classifiers are not stated as deposited or linked, and the QuPath release used for the analyses is not specified.
Limitations (5)
  • There is no consensus on clinically relevant PD-L1 epitopes or the best scoring system, and the clinical value of separating tumor-epithelial from inflammatory-cell PD-L1 staining is not yet established. · stated by the authors
  • Immunohistochemistry carries inherent staining artefacts that can mimic true expression, so classifiers had to exclude artefacts and necrosis, and patient scores used the median of up to three cores to limit artefact-driven outliers. · stated by the authors
  • Prognostic studies of p53 immunohistochemistry have been inconsistent, and the extreme-negative aberrant pattern scored here was described only recently and has rarely been evaluated in colorectal cancer cohorts. · stated by the authors
  • The text reports no ground-truth accuracy for cell detection, cell classification or stroma segmentation, no concordance with manual pathologist scores or other software, and no inter-operator test of manual steps such as grid correction, stain-vector choice and classifier training. · noted in this summary
  • Survival associations rest on unadjusted log-rank tests in a single cohort, split at cohort medians or pathologist-chosen p53 cutoffs, with no multivariable modeling or independent validation set, so prognostic independence from stage and other factors is untested. · noted in this summary

Read this if you score immunohistochemical biomarkers or tissue composition in whole slide images or TMAs and want an open-source, scriptable alternative to commercial pathology software with worked colon cancer survival examples.

Citations
8,887
FWCI
101.07
Field percentile
top 1%

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISPocock '21Grion '26Cicconet '21Zhang '25Madan '26Kendall '19
62% semantic matchBioengineering
J. Pocock, S. Graham, Q. D. Vu+11 more·bioRxiv·Dec 23, 2021