standOff. When we have changed our mind. Some entries in the table begin with a word in capitals and the reason for it. EVIDENCE CORRECTED — the finding still stands, but we had quoted the wrong sentence from the paper, and the quotation has been replaced with the right one. RETRACTED — the entry itself was wrong; it stays on the page, marked, with the reason, rather than being deleted. WITHDRAWN AND REPLACED — the entry was wrong, and a corrected version of it appears in the entries below. Each mark carries its own reason, in brackets, because the reasons differ from one entry to the next. An entry we have WITHDRAWN is removed from this page and kept in the record, with its reason. A line reading Audit and a date is a later re-reading of the entry against its whole sentence: DOUBTFUL means it was kept, and the reason says what is loose about it. Quotation marks inside the tagger's sentence. Where a tag is the tagger's own sentence, a phrase in quotation marks inside it is borrowed from the paper — the tagger's wording around it is not. The paper's own sentence is the evidence quote shown on the same row, so the borrowing can be checked against the thing borrowed from without leaving this page. Nothing inside quotation marks anywhere on this page is ours: if it is quoted, it is the paper's. Whose words a concept carries. A concept entry records a term the paper itself uses, so the term, the sense given for it, and its evidence are the paper's own words, quoted where the paper fixes them. This is the opposite of a sense entry in the tag table, which is the tagger's sentence about how the paper uses a term. Where a concept entry could not be grounded in the paper's words, the entry says so on its own row.| relation to other work | claim_relations | how this paper stands to another: agrees, contradicts, extends, qualifies |
| claim type | claim_type | the shape of the claim: existence, mechanism, method, critique, prediction, null |
| concept edge | concept_edges | an asserted relation between two concepts, with the sentence asserting it |
| concept | concepts | an idea the paper works with, quoted where the paper uses it |
| content tag | content_tags | a finding, definition, quantity, statistic or variable named in the paper |
| contribution | contribution | what the paper claims to add |
| control | controls | an experiment run to eliminate a rival explanation |
| limitation | limitations | what the paper says its evidence cannot support |
| mechanism | mechanisms | how the reported effect is said to come about |
| method | method | how the paper produced its evidence |
| object of study | object_of_study | what the paper is about — the thing it investigates |
| phenomenon | phenomenon | the effect or regularity under study |
| research question | research_question | the question the paper sets itself |
| sense | senses | a distinct use of a term, quoted where the paper fixes it |
| unrecordable relation | unhoused_relations | a relation the paper asserts that this vocabulary cannot express, with the reason |
QuPath: Open source software for digital pathology image analysis
Abstract
object_of_studycontributionclaim_typeconcepts↗concepts↗concepts↗concepts↗concepts↗concepts↗concepts↗concepts↗claim_relationscontent_tagsQuPath 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, QuPath’s flexible design makes it suitable for a wide range of additional image analysis applications across biomedical research.
Introduction
research_questionphenomenonlimitationslimitationsconcepts↗concepts↗concepts↗concepts↗concepts↗concepts↗concepts↗sensescontent_tags↗content_tagsThe ability to acquire high resolution digital scans of entire microscopic slides with high-resolution whole slide scanners is transforming tissue biomarker and companion diagnostic discovery through digital image analytics, automation, quantitation and objective screening of tissue samples. This area has become widely known as digital pathology 1 , 2 . Whole slide scanners can rapidly generate ultra-large 2D images or z-stacks in which each plane may contain up to 40 GB uncompressed data. Manual subjective scoring of this data by traditional pathologist assessment is no longer sufficient to support large-scale tissue biomarker trials, and cannot ensure the high quality, reproducible, objective analysis essential for reliable clinical correlation and candidate biomarker selection. New and powerful software tools are urgently required to ensure that pathological assessment of tissue is practical, accessible and reliable for biological discovery and the development of clinically-relevant tissue diagnostics.
limitationslimitationsconcepts↗concepts↗concepts↗concepts↗concepts↗concepts↗concepts↗concepts↗concepts↗concepts↗concepts↗concepts↗concepts↗concepts↗concept_edgesIn recent years, a vibrant ecosystem of open source bioimage analysis software has developed. Led by ImageJ 3 , researchers in multiple disciplines can now choose from a selection of powerful tools, such as Fiji 4 , Icy 5 , and CellProfiler 6 , to perform their image analyses. These open source packages encourage users to engage in further development and sharing of customized analysis solutions in the form of plugins, scripts, pipelines or workflows – enhancing the quality and reproducibility of research, particularly in the fields of microscopy and high content imaging. This template for open-source development of software has provided opportunities for image analysis to add considerably to translational research by enabling the development of the bespoke analytical methods required to address specific and emerging needs, which are often beyond the scope of existing commercial applications 7 . However, none of the aforementioned software applications tackle the specific visualization and computational challenges posed by whole slide images (WSI) and very large 2D data. Rather, open source tools for digital pathology to date have comprised libraries to handle digital slide formats (e.g. OpenSlide 8 , Bio-Formats 9 ), software to crop whole slide images into manageable tiles or perform analysis on such cropped tiles (e.g. SlideToolKit 10 , ImmunoRatio 11 ), or web platforms for data management and collaborative analysis (e.g. Cytomine 12 ). While each of this makes a valuable contribution, the field continues to lack a commonly-accepted, open software framework for developing and distributing novel digital pathology algorithms in a manner that is immediately accessible for any researcher or pathologist. In practice, this has meant that users without access to expensive commercial solutions have had to either resort to inefficient workarounds (such as image downsampling and cropping) to apply limited quantitative analysis using general open source analysis tools to a subset of their data 10 , 13 , or to rely primarily on laborious manual evaluation of slides, which is known to have high variability and limited reproducibility 14 , 15 . It has also made it more difficult for computational researchers to innovate in algorithm development, and to make state-of-the-art analysis methods widely available 16 .
QuPath ( https://qupath.github.io ) mechanisms↗concepts↗concepts↗concepts↗concepts↗concepts↗unhoused_relationsunhoused_relationssensessensesclaim_relationshas been developed to address these needs by offering the first comprehensive, open source desktop software application specifically designed to analyze and explore whole slide imaging data. At its core is a cross-platform, multithreaded, tile-based whole slide image viewer, which incorporates extensive annotation and visualization tools. On top of this, QuPath offers an array of novel algorithms to provide not only ready-made, user-friendly solutions to common, challenging analysis problems in pathology, but also the building blocks to create custom workflows – and link these together for batch processing with powerful scripting functionality ( Fig. 1 ) . Finally, QuPath enables developers to add their own extensions to solve new challenges and applications, and to exchange data in a streamlined manner with existing tools that otherwise provide limited whole slide support, such as ImageJ and MATLAB.
Illustration of QuPath’s use and functionality. ( a ) A typical workflow for TMA analysis (here, p53) demonstrates several of QuPath’s main features (left-to-right): Creation of a multi-slide project with automated TMA dearraying, stain estimation, cell detection and feature computation, trainable cell classification, batch processing, and survival analysis. ( b ) QuPath offers a wide range of additional functionality, including support for whole face tissue sections and fluorescence image analysis, data exchange with existing software and platforms (e.g. ImageJ and MATLAB), scriptable data mining, and rapid generation, visualization and export of spatial, morphological and intensity-based features.
A key feature underpinning QuPath’s mechanisms↗concepts↗concepts↗concepts↗concepts↗concepts↗concepts↗unhoused_relationssensessensesconcept_edgescontent_tagsfunctionality, and a major technical distinguishing factor between QuPath and other bioimaging analysis software, is its hierarchical, ‘object-based’ data model. Here, an ‘object’ refers primarily to a structure or region within the image, which may be created and manipulated by either interactive drawing tools (e.g. to annotate a particular region of interest) or automated segmentation commands (e.g. to detect individual nuclei or cells). However, in addition to representing a region of interest, objects can also be of different types (e.g. detection, annotation) and support the assignment of classifications, measurements and links to ‘parent’ and ‘child’ objects in a manner that can be rapidly queried and manipulated using built-in command or scripting. This generic model allows QuPath to represent and display relationships between very large numbers of image objects in an efficient and intuitive manner across gigapixel images, and support the fast and interactive training of object classifiers using machine learning techniques.
mechanisms↗concepts↗concepts↗concepts↗concepts↗concepts↗concepts↗sensesconcept_edgescontent_tagsA practical example of this is in the evaluation of the presence, localization and intensity of expression of key diagnostic, prognostic and predictive biomarkers in tissue sections. These biomarkers are typically detected using antibodies and chromogenic based detection systems, and are selectively expressed in tumor cells or in other cellular compartments. QuPath’s built-in cell segmentation algorithms can detect potentially millions of cells as objects within a single WSI, in addition to measuring cell morphology and biomarker expression. QuPath further supports the classification of different cell types according to these features, to generate a comprehensive phenotypic description of each cell within the tissue sample. This in turn provides a quantitative cellular map of the entire tissue section, which can be subsequently selected, queried and filtered to mine the image data and uncover morphological subtleties not immediately visible during traditional pathological assessment. All of this can typically be achieved within minutes, without a requirement for specialist hardware.
Results and Discussion
method↗concepts↗concepts↗concept_edgescontent_tagscontent_tagsTo demonstrate some of these capabilities, including its biological and clinical validity, we used QuPath to analyze several image sets derived from surgical resection specimens from a population-based cohort of 660 patients with stage II and stage III colon cancer, diagnosed between 2004–2008 (392 stage II, 268 stage III) and with high-quality curation of clinicopathological information. From representative paraffin-embedded tumor blocks, provided via the Northern Ireland Biobank, whole sections were haematoxylin and eosin-stained in the Northern Ireland Molecular Pathology Laboratory. Tissue microarrays (TMAs) were generated from representative 1 mm diameter tumor cores from each case, sampled in triplicate from the tumor center after annotation.
claim_relationsThe aim was to test the performance of QuPath in the context of known and novel biological and clinical data, using cancer immunology and tumor suppressor genes as paradigms. controlsconcepts↗concepts↗concepts↗concepts↗concepts↗concepts↗unhoused_relationsconcept_edgescontent_tagscontent_tagscontent_tagscontent_tagscontent_tagscontent_tagscontent_tagscontent_tagscontent_tagscontent_tags↗content_tags↗Firstly, we applied QuPath to analyze TMAs immunohistochemically-stained (IHC) for the T cell markers CD3 and CD8. Following digital scanning of the WSI, the initial QuPath setup required importing the images, applying automated ‘dearraying’ to identify tissue cores, manually refining the resulting dearrayed grid, and removing cores unsuitable for analysis. After this step, batch analysis was applied across all TMA slides to identify the tissue within each core and automatically count the number of positive cells per mm 2 tissue based upon a fast peak-finding algorithm after stain separation by color deconvolution 17 (Supplementary Fig. 1 , Supplementary Video 1 ). Running on a standard Mac Pro (3.5 GHz, 6-Core Intel Xeon E5, 32-GB RAM), this approach required less than 4 minutes for each biomarker to analyze 21 whole slide images, representing a total of >2000 tissue cores per biomarker and counting approximately 1.2 million CD3 positive and 0.6 million CD8 positive cells, while simultaneously exporting low-resolution images of each core both with and without markup for verification. After applying a median cutoff to the exported results, a statistically significant association between disease-specific survival and positive cell density scores was demonstrated for both CD3 and CD8 (log-rank test, p-values 0.006 and 0.007 respectively; Fig. 2a,b ). This recapitulates within our cohort the seminal work of Galon et al . in demonstrating the prognostic relevance of adaptive immunity in colorectal cancer 18 , while demonstrating a highly-efficient method of investigating similar markers within other cohorts and cancer types.
Survival analysis of colon cancer cohort based on QuPath automated image analysis. ( a – d ) Kaplan Meier survival analysis for biomarker scores of TMAs stained for CD3, CD8, p53 and PD-L1. Median cutoffs are applied in all cases, except p53 where two cutoffs were selected by an experienced pathologist to distinguish between aberrant negative, “wild type” (normal) and aberrant positive groups. Representative images showing an original core and QuPath markup image are included below. ( e ) Kaplan Meier curve showing patient stratification based on median tumor stromal percentage. Representative images show the original images and markup for tumors with a high and low stromal percentage respectively. Green indicates regions classified as stroma, dark red indicates tumor epithelium, while yellow represents other classified tissue or whitespace.
concepts↗concepts↗concepts↗concept_edgescontent_tagscontent_tagsNext, we used QuPath to evaluate immunohistochemistry for p53 in a second set of TMAs from the same cohort. This required a more sophisticated analysis to encompass the biological understanding and staining pattern of the marker. After applying QuPath’s cell detection algorithm to segment and measure cells within each core, a random trees classifier 19 was interactively trained to enable p53 expression to be scored selectively within the epithelial cell population according to nuclear staining intensity and proportion (Supplementary Fig. 2 , Supplementary Video 2 ). This analysis showed that aberrant p53 expression (diffuse intense or completely absent immunoreactivity) is associated with significantly poorer unadjusted disease-specific survival when compared with intermediate, ‘wild-type’ expression (p = 0.003 extreme negative/positive vs. intermediate; Fig. 2c ). Despite the well-established role of TP53 in colorectal cancer carcinogenesis, results from prognostic studies assessing p53 IHC expression have been inconsistent 20 . However, the extreme negative pattern of aberrant p53 immunoreactivity has only been described relatively recently 21 , 22 and has not been widely assessed in colorectal cancer cohorts. This example therefore emphasises the flexibility of the QuPath open source platform in measuring common IHC markers with variable tumor expression patterns, and demonstrates how the relationship between quantitative cellular analysis and clinical outcome can be robustly assessed.
limitationslimitationsconcepts↗concepts↗concepts↗concepts↗concept_edgesconcept_edgesWe then applied QuPath to the analysis of programmed cell death ligand 1 (PD-L1) immunoexpression in the same TMA cohort. PD-L1 immunoexpression is a prognostic marker in a range of cancer types and also a predictor of response to immune checkpoint therapy in some cancers 23 , 24 . However, there is a lack of consensus on the epitopes of clinical relevance and, more importantly, the optimal scoring systems for evaluation. The approach to analysis was similar in principle to that adopted for p53, however further attention was required because of additional challenges posed by PD-L1 immunostaining. Firstly, staining is cytoplasmic and/or membranous rather than nuclear. Secondly, PD-L1 can be expressed in tumor epithelium, but is more commonly expressed in other tissue compartments, notably within peritumoral stromal inflammatory cells. Although this level of heterogeneity in staining pattern is increasingly being recognised, the clinical importance of distinguishing tumor epithelial from inflammatory cell staining is yet to be fully understood. Sufficient cell classification is therefore required to identify both positively and negatively staining tumor (epithelial) and non-tumor cell populations, in addition to distinguishing true protein expression levels from the various staining artefacts that are inherent with IHC-based tissue analysis. concepts↗concepts↗sensesconcept_edgescontent_tags↗Here, applying QuPath, a cell was classified as positive or negative based on maximal DAB staining intensity, as a surrogate marker of protein expression, within a full cell region approximated by expanding detected nuclei (Supplementary Fig. 3 ). The percentage of all cells exhibiting PD-L1 positivity was calculated for each TMA core to give a summary PD-L1 score for that core. limitationsunhoused_relationscontent_tagscontent_tagscontent_tagscontent_tagsA median cutoff of 1.46% positive cells was determined from across the cohort to stratify patients, showing higher PD-L1 expression to be significantly associated with improved disease-specific survival in unadjusted analysis (p = 0.004, Fig. 2d ). Additionally, an analysis based on tertiles exhibited a dose-response effect, and separation of the tumor (epithelial) and non-epithelial components suggested that PD-L1 expression in colon cancer tissue is primarily found in the non-epithelial compartment (Supplementary Fig. 4 ). These results support the incipient evidence of PD-L1 prognostic value in colorectal cancer reported by ourselves and others in independent cohorts 24 – 26 , and may be of help when used together with tumor microsatellite instability status for patient stratification in consideration of anti-PD-L1 therapy.
Finally, to demonstrate QuPath’s concepts↗concepts↗concepts↗unhoused_relationsconcept_edgescontent_tagscontent_tagscontent_tagscontent_tags↗flexibility beyond IHC scoring in TMAs, we applied texture-based analysis to calculate the tumor stromal percentage in whole face representative tumor sections stained with hematoxylin and eosin (H&E) from 312 patients with stage II colon cancer from the same cohort. Several studies have reported that a higher intratumoral stromal percentage correlates with a worse prognosis in patients with stage II and stage III colorectal cancer 27 , 28 . Given highly variable clinical outcomes within stage II disease, there is a particular need for additional prognostic features in this group, to identify those patients most likely to benefit from adjuvant chemotherapy. These studies employed a crude visual estimate of tumor stromal percentage. Applying QuPath’s more fine-grained and reproducible assessment, and a median cutoff for analysis, we found lower disease-specific survival among patients with a high tumor stromal percentage in unadjusted analysis (p = 0.013; Fig. 2e ) , consistent with previous findings 27 , 28 .
limitationscontent_tagsIn summary, QuPath represents a uniquely comprehensive, user-friendly, open source bioimage analysis platform designed for whole slide images. We have demonstrated that it provides the tools necessary for fast, accurate and reproducible digital pathology analysis across a range of challenging applications. All of the above represent analyses that are cumbersome and time-consuming for pathologists to perform manually, given that they depend upon the accurate visual estimation of proportional staining within large numbers of stained cells, or of proportional composition of complex tissue areas. Accurate assessment is particularly difficult when the clinically relevant cutoff is very low and expression very localized (e.g. PD-L1), requiring detailed examination and precision. Such precise analyses are becoming increasingly necessary and important, both clinically in the era of personalized medicine, and in a research context for high-throughput evaluation of novel biomarkers. By offering an extensible environment for pathologists, biologists, and computer scientists to build highly performant algorithms for image interpretation and analysis, there is potential to drive adoption of quantitative imaging in academic, diagnostic and pharmaceutical research organizations, and to accelerate biomarker discovery in large scale multinational clinical trials. This is an absolute requirement to ensure a return on the global investment made in companion diagnostics for precision medicine.
Materials and Methods
Cohort
This analysis was based on a population-based cohort of 660 stage II/III colon adenocarcinoma patients, representing 89% of all patients undergoing surgery for stage II/III colon adenocarcinoma in two Healthcare Trusts in Northern Ireland between 2004 and 2008. Patients were sampled and clinical information obtained from the Northern Ireland Cancer Registry. All patients were followed up for occurrence and cause of death via linkage to the Northern Ireland Registrar General’s Office up to 31 st December 2013. Ethical approval for these linkages was given by ORECNI (REC: 10/NIR02/53). Over ten years (and a mean of 5 years) of clinical follow-up, 46% of patients had died, of which 38% were from CRC-specific causes. Corresponding tumor slides and blocks were retrieved and collated via the Northern Ireland Biobank which has ethical approval to use de-identified tissue samples from the Belfast Health and Social Care Tissue Pathology archive (REC:11/NI/0013).
This study was conducted in accordance to REporting recommendations for tumour MARKer prognostic studies (REMARK) 29 . The reporting standards of the current study fulfill these recommendations.
Following slide review, a new section was cut for H&E staining from a single representative tumor block in each case, and the new slides annotated for tissue microarray (TMA) construction. Three representative areas within the tumor center of each block were annotated for targeted coring (by an experienced biomedical scientist and confirmed by expert pathologists, MBL and JAJ). One millimeter diameter tissue cores were extracted from donor blocks and inserted into recipient blocks using a manual tissue arrayer (Estigen, Tartu, Estonia).
Immunohistochemistry
All IHC was performed in a hybrid laboratory (Northern Ireland Molecular Pathology Laboratory) that has UK Clinical Pathology Accreditation. Internally validated biomarker conditions, which followed UK-NEQAS guidelines (CD3 and CD8) or were based on expected performance from the literature (p53 and PD-L1), were as follows: CD3 (clone 2GV6 Ventana BenchMark; CC1 32 minutes, Optiview detection), CD8 (clone c8/144B, Dako: Leica Bond III, ER2 20 mins, 1/50, polymer detection), p53 (clone DO-7, Dako, ER2 30 mins, 1/100, polymer detection), PD-L1 (clone SP142, Ventana BenchMark, CC1 24 mins, optiview detection).
Image data acquisition
limitationsAll TMA slides were scanned using an Aperio ScanScope CS whole slide scanner at 40X magnification, with a resolution of 0.25 μm/pixel. H&E slide scanning was heterogenous: 231 were scanned on the Aperio ScanScope scanner, while 81 were scanned on a Hamamatsu Nanozoomer. The resolution of all images was within the range 0.231–0.253 μm/pixel.
Software
unhoused_relationsunhoused_relationsclaim_relationsQuPath was written as a novel, cross-platform Java application. The core software was developed using Java 8, with a user interface written using JavaFX. Whole slide image reading made use of QuPath’s interfaces to the OpenSlide library 8 , while the implementations of QuPath’s cell detection and superpixel segmentation commands made use of ImageJ as a library for standard image processing operations 3 . The random trees classifier and fast cell counting were implemented using the OpenCV library ( http://opencv.org ). The analysis for this study was performed on a Mac Pro (3.5 GHz, 6-Core Intel Xeon E5, 32-GB RAM).
Code availability
Source code and documentation for QuPath are available at https://qupath.github.io .
Statistical analysis
mechanisms↗controlsconcepts↗concepts↗concepts↗concepts↗concepts↗unhoused_relationssensesconcept_edgescontent_tags↗Survival curves using the Kaplan Meier method were generated and log-rank tests applied using the TMA data viewer within QuPath, and independently verified using R (version 3.2.2) 30 with the ‘Survival’ package (version 2.38–3) 31 . For the calculation of disease-specific survival, deaths from other causes were treated as censored events. Median cutoff values were used in all cases, except for p53 where an experienced pathologist (MBL) selected two biologically-plausible cutoffs (H-scores 10 and 160) to separate extreme positive and extreme negative cases from those with intermediate (‘wild-type’) expression, based upon viewing all TMA cores post-analysis ranked by H-score. Stratification based on tertiles is also provided in the Supplementary Materials for PD-L1.
For TMA analysis, up to three tissue cores were available from each tumor, all selected from the same paraffin block representing the central tumor region. A single patient biomarker score was defined as the median of all available scores for the corresponding patient and biomarker. The median was chosen to aid the robustness of the measurement in a high-throughput setting, and reduce the likelihood of basing the score for any individual patient on an outlier that may have been caused by a tissue or staining artefact.
Tissue microarray preprocessing
mechanisms↗mechanisms↗controlsconcepts↗concepts↗concept_edgesSeparate projects were created within QuPath for each biomarker, and the slide images imported to the corresponding projects. QuPath’s automated TMA dearrayer was applied in batch over all slides within each project to identify tissue cores. The resulting TMA grid was manually verified and amended where necessary, e.g. to adjust the locations of cores that were outside their expected position, or to remove cores where prominent artefacts were visible. Patient identifiers were then imported into QuPath for each core to assist alignment with survival data later. Additionally, stain vector (i.e. color) and background estimates were applied for each IHC analysis project to improve stain separation within QuPath using color deconvolution 17 . This was achieved by selecting a representative area containing an area of background along with examples of strong hematoxylin and DAB staining, and applying QuPath’s Estimate stain vectors command to identify stain vectors within this region. The resulting vectors were then used for all images in the project.
Analysis of CD3 and CD8 IHC
mechanisms↗After the initial TMA preprocessing steps described above, analysis of CD3 and CD8 was performed using QuPath’s Simple tissue detection and Fast cell counts commands. Briefly, tissue was detected within each TMA core by thresholding a downsampled and smoothed image of the core, followed by cleanup of the resulting binary image by morphological operations. Individual cells were identified by separating stains using color deconvolution and identifying peaks in either the hematoxylin channel (CD3) or sum of the hematoxylin and DAB channels (CD8) after smoothing, and assigning these as positive or negative cells based upon the smoothed DAB channel information. The number of positive cells and area detected were used to calculate the average number of positive cells per mm 2 , and these results exported along with markup images showing the detected cells, for visual verification. The detection and export steps were fully automated using a batch processing script.
Analysis of p53 IHC
Preprocessing steps were applied as described above. QuPath’s mechanisms↗concepts↗concepts↗concepts↗sensesconcept_edgescontent_tags↗content_tagscontent_tags↗content_tags↗content_tags↗Cell detection command was then used to identify cells across all cores based upon nuclear staining. This command additionally estimates the full extent of each cell based upon a constrained expansion of the nucleus region, and calculates up to 33 measurements of intensity and morphology, including nucleus area, circularity, staining intensity for hematoxylin and DAB, and nucleus/cell area ratio. Because not all of these measurements are expected to provide independent or useful information with regard to cell classification, a subset of 16 measurements was chosen empirically and supplemented for each cell by measuring the local density of cells, and taking a Gaussian-weighted sum of the corresponding measurements within neighboring cells using QuPath’s Add smoothed features command. A two-way random trees classifier was then interactively trained to distinguish tumor epithelial cells from all other detections (comprising non-epithelial cells, necrosis, or any artefacts misidentified as cells) and applied across all slides (see Supplementary Video 2 ). Intensity thresholds were set to further subclassify tumor cells as being negative, weak, moderate or strongly positive for p53 staining based upon mean nuclear DAB optical densities. An H-score was calculated for each tissue core by adding 3x% strongly stained tumor nuclei, 2x% moderately stained tumor nuclei, and 1x% weakly stained tumor nuclei 32 , giving results in the range 0 (all tumor nuclei negative) to 300 (all tumor nuclei strongly positive).
Analysis of PD-L1 IHC
mechanisms↗content_tags↗content_tags↗The approach to scoring PD-L1 was similar to p53, apart from the following: 1) a three-way random trees classifier was trained to distinguish between epithelial, non-epithelial and ‘other’ detections (including artefacts and necrosis); 2) cells were classified as positive or negative based upon a single intensity threshold applied to the maximum DAB optical density within the full cell area, and 3) summary scores were generated as the percentage of cells classified as positive, with ‘other’ detections removed.
Analysis of tumor stromal percentage in H&E whole face sections
mechanisms↗concepts↗concepts↗concepts↗content_tags↗content_tags↗Representative tumor regions were annotated across all 312 H&E-stained slides by an experienced pathologist (MBL) using QuPath’s manual annotation tools. A script was then applied in batch to automatically identify and set the average background intensity for the red, green and blue channels of each image, which varied markedly according to the scanner used. A second script was then run over all images to apply QuPath’s SLIC superpixel segmentation command to subdivide each annotated region into ‘superpixels’ based upon simple linear iterative clustering 33 . This script additionally calculated both the average hue for each superpixel along with Haralick texture features 34 from optical density values using QuPath’s Add intensity features command. QuPath’s Add smoothed features command was also applied to calculate a Gaussian-weighted sum of the features of neighboring superpixels, and append these to the existing features for each superpixel. This provide additional contextual information extending beyond the superpixel itself.
mechanisms↗concepts↗concepts↗concepts↗concepts↗sensesconcept_edgesconcept_edgesconcept_edgescontent_tags↗content_tags↗content_tagsA subset of 40 ‘training’ images was then identified for the pathologist to interactively train a random trees classifier to distinguish between tissue areas comprising tumor epithelium, stroma and ‘other’ (e.g. whitespace, mucin, normal muscle or necrosis). This required drawing around regions containing tissue of each class and annotating these accordingly. During this process, QuPath used all available features to train the classifier in a background process and thereby provide immediate feedback on classification performance. Once the classification was considered adequate across the training images, the classifier was applied to all images within the set and the total area of superpixels for each class was exported. The tumor stromal percentage (TSP) was then calculated as
where AS represents the total area classified as stroma, and AE represents the total area classified as epithelium.
Electronic supplementary material
Supplementary Information
Supplementary Video 1
Supplementary Video 2
Acknowledgements
The research leading to these results has received funding from Invest Northern Ireland (RDO0712612 to P.W.H.); Cancer Research UK Accelerator (C11512/A20256 to P.W.H./M.S.-T./J.A.J). The Northern Ireland Molecular Pathology Laboratory is supported by Cancer Research UK, Experimental Cancer Medicine Centre Network, the NI Health and Social Care Research and Development Division, the Sean Crummey Memorial Fund, the Tom Simms Memorial Fund and the Friends of the Cancer Centre (to M.S.-T.). The Northern Ireland Biobank is funded by the Health and Social Care Research and Development Division of the Public Health Agency in Northern Ireland and Cancer Research UK (to J.A.J/P.W.H/S.McQ ref. SPI/5069/14; additional support was received from the Friends of the Cancer Centre (to J.A.J./P.W.H/M.B.L). The cohort was supported by Cancer Research UK (ref. C37703/A15333 and ref. C50104/A17592) and the HSC R&D office of Northern Ireland (ref. EAT/4905/13). The Northern Ireland Cancer Registry is funded by the NI Public Health Agency. We thank Mr Ken Arthur for the construction of the original tissue microarrays used in this study, Ms Victoria Bingham for her work in staining and scanning the slides, and both Dr Roisin O’Neill and the Northern Ireland Cancer Registry for their contributions to the clinical data collation.
Author Contributions
P.B. designed and developed the software. Feedback and testing of the software was provided by J.A.F., Y.D., D.G.M., P.W.H., P.D.D. and M.B.L. Before scanning and image analysis the expected expression of the biomarkers on tissue sections was initially assessed microscopically as fit-for-purpose by a team of experienced histopathologists M.S.-T., S.Mc.Q., J.A.J. Experiments were designed by P.B., M.B.L., M.S.-T. and P.W.H. and performed by P.B. and M.B.L. Data was analyzed by P.B., J.A.F. and H.G.C. Manuscript was written by PB with contributions from all authors.
Competing Interests
The authors declare that they have no competing interests.
Footnotes
Contributor Information
Manuel Salto-Tellez, Email: m.salto-tellez@qub.ac.uk.
Peter W. Hamilton, Email: p.hamilton@qub.ac.uk
References
Associated Data
Supplementary Materials
Supplementary Information
Supplementary Video 1
Supplementary Video 2
Relations on this page carry a quotation for the claim. Where a relation was separately evidenced — and so tested against the vocabulary for its type — the row says so; 16 of 16 relations here carry that mark. The rest are the tagger’s reading of the sentence quoted beside them.
| facet | tag | field | conf. | whose words |
|---|---|---|---|---|
| object_of_study | QuPath's architecture, whole-slide analysis workflows and biomarker-analysis validity | object_of_study | the paper's words | |
| research_question | whether an open, extensible desktop platform can make whole-slide digital pathology fast, accessible, quantitative and reproducible | research_question | the paper's words | |
| contribution | QuPath, an extensible whole-slide platform with object-based analysis, scripting and validated biomarker workflows | contribution | the paper's words | |
| claim_type | method | claim_type | the paper's words | |
| phenomenon | computational and reproducibility challenges in digital pathology and tissue-biomarker scoring | phenomenon | the paper's words | |
| method | system-or-tool-building also in: system-or-tool-building | method | the paper's words | |
| mechanisms | A cross-platform multithreaded tile viewer supports navigation and analysis of gigapixel slides. also in: tile-based whole-slide viewing | mechanisms[0] | 0.95 | the tagger's words |
| mechanisms | Objects carry regions, types, classes, measurements and parent-child links that scripts can query. also in: hierarchical object representation | mechanisms[1] | 0.95 | the tagger's words |
| mechanisms | Cell segmentation, morphology and expression measurements feed cell-type classification and a quantitative cellular map. also in: cellular phenotyping | mechanisms[2] | 0.95 | the tagger's words |
| mechanisms | A batch dearrayer locates tissue cores before manual grid verification and removal of unsuitable cores. also in: automated TMA dearraying | mechanisms[3] | 0.95 | the tagger's words |
| mechanisms | Representative background, hematoxylin and DAB regions estimate stain vectors used across a project. also in: stain-vector estimation | mechanisms[4] | 0.95 | the tagger's words |
| mechanisms | Color deconvolution and peak finding classify cells and calculate positive cells per square millimeter. also in: peak-based positive-cell counting | mechanisms[5] | 0.95 | the tagger's words |
| mechanisms | A random-trees model separates epithelial from other detections before intensity thresholds assign p53 classes. also in: p53 cell classification | mechanisms[6] | 0.95 | the tagger's words |
| mechanisms | Maximum DAB optical density within an expanded cell region assigns positive or negative status. also in: PD-L1 positivity threshold | mechanisms[7] | 0.95 | the tagger's words |
| mechanisms | SLIC superpixels receive hue, Haralick and neighborhood-smoothed features for tissue classification. also in: superpixel texture representation | mechanisms[8] | 0.95 | the tagger's words |
| mechanisms | Pathologist annotations train a background random-trees classifier with immediate feedback before batch application. also in: interactive tissue classifier | mechanisms[9] | 0.95 | the tagger's words |
| mechanisms | The median of available core scores reduces dependence on a tissue or staining outlier. also in: robust patient aggregation | mechanisms[10] | 0.95 | the tagger's words |
| limitations | A single uncompressed whole-slide plane may contain up to 40 GB of data. | limitations[0] | 0.95 | the tagger's words |
| limitations | Manual subjective scoring is inadequate for large trials and cannot ensure reproducible objective analysis. | limitations[1] | 0.95 | the tagger's words |
| limitations | Earlier open-source tools did not tackle the visualization and computation demands of whole-slide images. | limitations[2] | 0.95 | the tagger's words |
| limitations | Manual slide evaluation has high variability and limited reproducibility. | limitations[3] | 0.95 | the tagger's words |
| limitations | PD-L1 lacks consensus on clinically relevant epitopes and optimal scoring systems. | limitations[4] | 0.95 | the tagger's words |
| limitations | PD-L1 heterogeneity and staining artefacts complicate compartment-specific positivity classification. | limitations[5] | 0.95 | the tagger's words |
| limitations | Precise manual assessment is difficult when a clinically relevant cutoff is very low and expression is localized. | limitations[6] | 0.95 | the tagger's words |
| limitations | H&E slides were acquired on two scanner models with heterogeneous resolutions. | limitations[7] | 0.95 | the tagger's words |
| limitations | The survival associations reported in the demonstrations are unadjusted analyses. | limitations[8] | 0.95 | the tagger's words |
| controls | exported low-resolution core images with and without cell markup eliminates: undetected cell-count or core-localization errors | controls[0] | 0.95 | the tagger's words |
| controls | independent R reproduction of Kaplan-Meier curves and log-rank tests eliminates: QuPath TMA-viewer implementation errors · outcome: held | controls[1] | 0.95 | the tagger's words |
| controls | manual verification and amendment of the automated TMA grid eliminates: mislocated or visibly artefactual tissue cores | controls[2] | 0.95 | the tagger's words |
| concepts | QuPath also in: QuPath | concepts[0] | 0.95 | the paper's words |
| concepts | bioimage analysis software also in: bioimage analysis software | concepts[1] | 0.95 | the paper's words |
| concepts | digital pathology also in: digital pathology | concepts[2] | 0.95 | the paper's words |
| concepts | whole slide image analysis also in: whole slide image analysis | concepts[3] | 0.95 | the paper's words |
| concepts | tumor identification also in: tumor identification | concepts[4] | 0.95 | the paper's words |
| concepts | high-throughput biomarker evaluation also in: high-throughput biomarker evaluation | concepts[5] | 0.95 | the paper's words |
| concepts | batch-processing also in: batch-processing | concepts[6] | 0.95 | the paper's words |
| concepts | scripting functionality also in: scripting functionality | concepts[7] | 0.95 | the paper's words |
| concepts | extensible platform also in: extensible platform | concepts[8] | 0.95 | the paper's words |
| concepts | whole slide scanners also in: whole slide scanners | concepts[9] | 0.95 | the paper's words |
| concepts | ultra-large 2D images also in: ultra-large 2D images | concepts[10] | 0.95 | the paper's words |
| concepts | traditional pathologist assessment also in: traditional pathologist assessment | concepts[11] | 0.95 | the paper's words |
| concepts | objective analysis also in: objective analysis | concepts[12] | 0.95 | the paper's words |
| concepts | candidate biomarker selection also in: candidate biomarker selection | concepts[13] | 0.95 | the paper's words |
| concepts | digital image analytics, automation, quantitation and objective screening of tissue samples | concepts[14] | 0.95 | the paper's words |
| concepts | open source bioimage analysis software | concepts[15] | 0.95 | the paper's words |
| concepts | open source packages also in: open source packages | concepts[16] | 0.95 | the paper's words |
| concepts | quality and reproducibility of research | concepts[17] | 0.95 | the paper's words |
| concepts | ImageJ also in: ImageJ | concepts[18] | 0.95 | the paper's words |
| concepts | Fiji also in: Fiji | concepts[19] | 0.95 | the paper's words |
| concepts | Icy also in: Icy | concepts[20] | 0.95 | the paper's words |
| concepts | CellProfiler also in: CellProfiler | concepts[21] | 0.95 | the paper's words |
| concepts | plugins also in: plugins | concepts[22] | 0.95 | the paper's words |
| concepts | pipelines also in: pipelines | concepts[23] | 0.95 | the paper's words |
| concepts | workflows also in: workflows | concepts[24] | 0.95 | the paper's words |
| concepts | whole slide images also in: whole slide images | concepts[25] | 0.95 | the paper's words |
| concepts | other bioimaging analysis software also in: other bioimaging analysis software | concepts[26] | 0.95 | the paper's words |
| concepts | OpenSlide also in: OpenSlide | concepts[27] | 0.95 | the paper's words |
| concepts | Bio-Formats also in: Bio-Formats | concepts[28] | 0.95 | the paper's words |
| concepts | Cytomine also in: Cytomine | concepts[29] | 0.95 | the paper's words |
| concepts | tile-based whole slide image viewer also in: tile-based whole slide image viewer | concepts[30] | 0.95 | the paper's words |
| concepts | annotation and visualization tools also in: annotation and visualization tools | concepts[31] | 0.95 | the paper's words |
| concepts | custom workflows also in: custom workflows | concepts[32] | 0.95 | the paper's words |
| concepts | batch processing also in: batch processing | concepts[33] | 0.95 | the paper's words |
| concepts | extensions also in: extensions | concepts[34] | 0.95 | the paper's words |
| concepts | object-based also in: object-based | concepts[35] | 0.95 | the paper's words |
| concepts | object also in: object | concepts[36] | 0.95 | the paper's words |
| concepts | region of interest also in: region of interest | concepts[37] | 0.95 | the paper's words |
| concepts | object classifiers also in: object classifiers | concepts[38] | 0.95 | the paper's words |
| concepts | machine learning techniques also in: machine learning techniques | concepts[39] | 0.95 | the paper's words |
| concepts | cell segmentation algorithms also in: cell segmentation algorithms | concepts[40] | 0.95 | the paper's words |
| concepts | cell morphology also in: cell morphology | concepts[41] | 0.95 | the paper's words |
| concepts | biomarker expression also in: biomarker expression | concepts[42] | 0.95 | the paper's words |
| concepts | classification of different cell types | concepts[43] | 0.95 | the paper's words |
| concepts | comprehensive phenotypic description also in: comprehensive phenotypic description | concepts[44] | 0.95 | the paper's words |
| concepts | quantitative cellular map also in: quantitative cellular map | concepts[45] | 0.95 | the paper's words |
| concepts | tissue microarrays also in: tissue microarrays | concepts[46] | 0.95 | the paper's words |
| concepts | TMAs also in: TMAs | concepts[47] | 0.95 | the paper's words |
| concepts | immunohistochemically-stained also in: immunohistochemically-stained | concepts[48] | 0.95 | the paper's words |
| concepts | IHC also in: IHC | concepts[49] | 0.95 | the paper's words |
| concepts | color deconvolution also in: color deconvolution | concepts[50] | 0.95 | the paper's words |
| concepts | peak-finding algorithm also in: peak-finding algorithm | concepts[51] | 0.95 | the paper's words |
| concepts | positive cell density scores also in: positive cell density scores | concepts[52] | 0.95 | the paper's words |
| concepts | disease-specific survival also in: disease-specific survival | concepts[53] | 0.95 | the paper's words |
| concepts | random trees classifier also in: random trees classifier | concepts[54] | 0.95 | the paper's words |
| concepts | aberrant p53 expression also in: aberrant p53 expression | concepts[55] | 0.95 | the paper's words |
| concepts | wild-type also in: wild-type | concepts[56] | 0.95 | the paper's words |
| concepts | programmed cell death ligand 1 also in: programmed cell death ligand 1 | concepts[57] | 0.95 | the paper's words |
| concepts | PD-L1 also in: PD-L1 | concepts[58] | 0.95 | the paper's words |
| concepts | tumor epithelial also in: tumor epithelial | concepts[59] | 0.95 | the paper's words |
| concepts | inflammatory cell staining also in: inflammatory cell staining | concepts[60] | 0.95 | the paper's words |
| concepts | PD-L1 score also in: PD-L1 score | concepts[61] | 0.95 | the paper's words |
| concepts | percentage of all cells exhibiting PD-L1 positivity | concepts[62] | 0.95 | the paper's words |
| concepts | hematoxylin and eosin also in: hematoxylin and eosin | concepts[63] | 0.95 | the paper's words |
| concepts | H&E also in: H&E | concepts[64] | 0.95 | the paper's words |
| concepts | tumor stromal percentage also in: tumor stromal percentage | concepts[65] | 0.95 | the paper's words |
| concepts | Kaplan Meier method also in: Kaplan Meier method | concepts[66] | 0.95 | the paper's words |
| concepts | log-rank tests also in: log-rank tests | concepts[67] | 0.95 | the paper's words |
| concepts | single patient biomarker score also in: single patient biomarker score | concepts[68] | 0.95 | the paper's words |
| concepts | median of all available scores also in: median of all available scores | concepts[69] | 0.95 | the paper's words |
| concepts | background estimates also in: background estimates | concepts[70] | 0.95 | the paper's words |
| concepts | stain separation also in: stain separation | concepts[71] | 0.95 | the paper's words |
| concepts | H-score also in: H-score | concepts[72] | 0.95 | the paper's words |
| concepts | tumor epithelial cells also in: tumor epithelial cells | concepts[73] | 0.95 | the paper's words |
| concepts | all other detections also in: all other detections | concepts[74] | 0.95 | the paper's words |
| concepts | SLIC superpixel segmentation also in: SLIC superpixel segmentation | concepts[75] | 0.95 | the paper's words |
| concepts | simple linear iterative clustering also in: simple linear iterative clustering | concepts[76] | 0.95 | the paper's words |
| concepts | Haralick texture features also in: Haralick texture features | concepts[77] | 0.95 | the paper's words |
| concepts | tissue areas also in: tissue areas | concepts[78] | 0.95 | the paper's words |
| concepts | tumor epithelium also in: tumor epithelium | concepts[79] | 0.95 | the paper's words |
| concepts | stroma also in: stroma | concepts[80] | 0.95 | the paper's words |
| concepts | TSP also in: TSP | concepts[81] | 0.95 | the paper's words |
| concepts | outlier also in: outlier | concepts[82] | 0.95 | the paper's words |
| unhoused_relations | QuPath → extensions QuPath enables developers to add extensions for new challenges and applications. | unhoused_relations[0] | 0.95 | the tagger's words |
| unhoused_relations | QuPath → ImageJ QuPath exchanges data with ImageJ despite ImageJ's limited whole-slide support. | unhoused_relations[1] | 0.95 | the tagger's words |
| unhoused_relations | object-based → object classifiers The object-based model supports fast interactive training of object classifiers. | unhoused_relations[2] | 0.95 | the tagger's words |
| unhoused_relations | positive cell density scores → disease-specific survival Positive CD3 and CD8 cell-density scores are significantly associated with disease-specific survival. | unhoused_relations[3] | 0.95 | the tagger's words |
| unhoused_relations | PD-L1 → disease-specific survival Higher PD-L1 expression is significantly associated with improved disease-specific survival. | unhoused_relations[4] | 0.95 | the tagger's words |
| unhoused_relations | tumor stromal percentage → disease-specific survival High tumor stromal percentage correlates with worse prognosis and lower disease-specific survival. | unhoused_relations[5] | 0.95 | the tagger's words |
| unhoused_relations | median of all available scores → outlier Median aggregation reduces the likelihood that an outlier determines a patient's biomarker score. | unhoused_relations[6] | 0.95 | the tagger's words |
| unhoused_relations | QuPath → OpenSlide QuPath uses OpenSlide as a library for whole-slide image reading. | unhoused_relations[7] | 0.95 | the tagger's words |
| unhoused_relations | QuPath → ImageJ QuPath uses ImageJ as a library for standard image-processing operations. | unhoused_relations[8] | 0.95 | the tagger's words |
| senses | digital pathology as digital analytics, automation, quantitation and objective screening of tissue samples | senses[0] | 0.95 | the tagger's words |
| senses | QuPath as an open, extensible whole-slide viewer, analysis environment and workflow platform | senses[1] | 0.95 | the tagger's words |
| senses | QuPath object as an interactively or automatically created image structure or region with type, class and measurements | senses[2] | 0.95 | the tagger's words |
| senses | object-based data model as a hierarchy connecting image objects through parent-child links | senses[3] | 0.95 | the tagger's words |
| senses | quantitative cellular map as the queryable phenotypic output of segmentation, measurement and classification | senses[4] | 0.95 | the tagger's words |
| senses | PD-L1 score as the percentage of all cells in a core classified as positive | senses[5] | 0.95 | the tagger's words |
| senses | H-score as a weighted sum of strong, moderate and weak p53-positive tumor nuclei | senses[6] | 0.95 | the tagger's words |
| senses | patient biomarker score as the median of available core scores | senses[7] | 0.95 | the tagger's words |
| senses | tumor stromal percentage as stromal area relative to stromal plus epithelial area | senses[8] | 0.95 | the tagger's words |
| senses | whole-slide analysis as tile-based viewing plus object detection, measurement, classification and batch scripting | senses[9] | 0.95 | the tagger's words |
| claim_relations | extends → open-source bioimage analysis software note: QuPath adds whole-slide, tumor-identification, biomarker and scripting capabilities. | claim_relations[0] | 0.95 | the tagger's words |
| claim_relations | addresses → whole-slide visualization and computational challenges note: The platform is explicitly developed to address the documented whole-slide gap. | claim_relations[1] | 0.95 | the tagger's words |
| claim_relations | builds-on → OpenSlide, ImageJ and OpenCV libraries note: The implementation section names each library and its role. | claim_relations[2] | 0.95 | the tagger's words |
| claim_relations | validates → known and novel biological and clinical data note: The case studies test performance against established and exploratory biomarker patterns. | claim_relations[3] | 0.95 | the tagger's words |
| concept_edges | tissue microarrays —same-as→ TMAs the relation itself is separately evidenced · strength of the relation: 1 · note: Direct acronym apposition. · the relation stated: Tissue microarrays are parenthetically abbreviated as TMAs. also in: tissue microarrays · TMAs | concept_edges[0] | 0.95 | the tagger's words |
| concept_edges | immunohistochemically-stained —same-as→ IHC the relation itself is separately evidenced · strength of the relation: 1 · note: Direct acronym apposition. · the relation stated: Immunohistochemically-stained is parenthetically abbreviated as IHC. also in: immunohistochemically-stained · IHC | concept_edges[1] | 0.95 | the tagger's words |
| concept_edges | programmed cell death ligand 1 —same-as→ PD-L1 the relation itself is separately evidenced · strength of the relation: 1 · note: Direct acronym apposition. · the relation stated: Programmed cell death ligand 1 is parenthetically abbreviated as PD-L1. also in: programmed cell death ligand 1 · PD-L1 | concept_edges[2] | 0.95 | the tagger's words |
| concept_edges | hematoxylin and eosin —same-as→ H&E the relation itself is separately evidenced · strength of the relation: 1 · note: Direct acronym apposition. · the relation stated: Hematoxylin and eosin is parenthetically abbreviated as H&E. also in: hematoxylin and eosin · H&E | concept_edges[3] | 0.95 | the tagger's words |
| concept_edges | tumor stromal percentage —same-as→ TSP the relation itself is separately evidenced · strength of the relation: 1 · note: Direct acronym apposition. · the relation stated: Tumor stromal percentage is parenthetically abbreviated as TSP. also in: tumor stromal percentage · TSP | concept_edges[4] | 0.95 | the tagger's words |
| concept_edges | QuPath —contrasted-with→ other bioimaging analysis software the relation itself is separately evidenced · strength of the relation: 0.95 · note: Direct unhedged technical distinction. · the relation stated: The hierarchical object model is stated as a major technical distinguishing factor between QuPath and other bioimaging software. also in: QuPath · other bioimaging analysis software | concept_edges[5] | 0.95 | the tagger's words |
| concept_edges | aberrant p53 expression —distinct-from→ wild-type the relation itself is separately evidenced · strength of the relation: 0.95 · note: Direct reported comparison of expression patterns. · the relation stated: Diffuse intense or absent aberrant p53 expression is compared with intermediate wild-type expression. also in: aberrant p53 expression · wild-type | concept_edges[6] | 0.95 | the tagger's words |
| concept_edges | tumor epithelial —distinct-from→ inflammatory cell staining the relation itself is separately evidenced · strength of the relation: 0.7 · note: The categories are explicitly distinguished, but their clinical importance is not yet understood. · the relation stated: The paper distinguishes tumor epithelial from inflammatory-cell PD-L1 staining while noting unresolved clinical importance.
Audit 2026-09-28: DOUBTFUL - read against its full span and kept, but the importance of distinguishing the two stainings 'is yet to be fully understood'; distinct-from is loose. also in: tumor epithelial · inflammatory cell staining | concept_edges[7] | 0.95 | the tagger's words |
| concept_edges | tumor epithelial cells —distinct-from→ all other detections the relation itself is separately evidenced · strength of the relation: 0.95 · note: Direct method specification. · the relation stated: A trained classifier distinguishes tumor epithelial cells from all other detections. also in: tumor epithelial cells · all other detections | concept_edges[8] | 0.95 | the tagger's words |
| concept_edges | PD-L1 score —operationalised-as→ percentage of all cells exhibiting PD-L1 positivity the relation itself is separately evidenced · strength of the relation: 1 · note: The paper states the calculation directly. · the relation stated: A core's summary PD-L1 score is calculated from the percentage of all cells exhibiting positivity. | concept_edges[9] | 0.95 | the tagger's words |
| concept_edges | single patient biomarker score —operationalised-as→ median of all available scores the relation itself is separately evidenced · strength of the relation: 1 · note: Exact operational definition. · the relation stated: A single patient biomarker score is defined as the median of available core scores. | concept_edges[10] | 0.95 | the tagger's words |
| concept_edges | tumor epithelium —part-of→ tissue areas the relation itself is separately evidenced · strength of the relation: 1 · note: Direct categorical containment statement. · the relation stated: The classifier distinguishes tissue areas comprising tumor epithelium, stroma and other tissue. also in: tumor epithelium · tissue areas | concept_edges[11] | 0.95 | the tagger's words |
| concept_edges | stroma —part-of→ tissue areas the relation itself is separately evidenced · strength of the relation: 1 · note: Direct categorical containment statement. · the relation stated: The classifier distinguishes tissue areas comprising tumor epithelium, stroma and other tissue. also in: stroma · tissue areas | concept_edges[12] | 0.95 | the tagger's words |
| concept_edges | open source packages —causes→ quality and reproducibility of research direction on the measure: increases · the relation itself is separately evidenced · strength of the relation: 0.9 · note: Direct general mechanism claim, stated without a modal but not as a necessary or sufficient condition. · the relation stated: The paper says open-source packages encourage shared customization, enhancing research quality and reproducibility. | concept_edges[13] | 0.95 | the tagger's words |
| concept_edges | background estimates —causes→ stain separation direction on the measure: increases · the relation itself is separately evidenced · strength of the relation: 0.6 — the record marks this weak · note: Purpose language supports a proposed improvement rather than a measured effect. · the relation stated: Background estimates are applied with stain vectors to improve stain separation. also in: background estimates · stain separation | concept_edges[14] | 0.95 | the tagger's words |
| concept_edges | classification of different cell types —causes→ comprehensive phenotypic description direction on the measure: increases · the relation itself is separately evidenced · strength of the relation: 0.7 · note: The paper presents this as the intended output of the classification step. · the relation stated: Cell-type classification is used to generate a comprehensive phenotypic description. | concept_edges[15] | 0.95 | the tagger's words |
| content_tags | QuPath combines whole-slide viewing, tumor identification, biomarker analysis, scripting and extension support. | content_tags.findings[0] | 0.95 | the tagger's words |
| content_tags | The hierarchical object model supports efficient relationships and interactive classifier training across gigapixel images. | content_tags.findings[1] | 0.95 | the tagger's words |
| content_tags | A whole slide can contain millions of detected, measured and classified cell objects. | content_tags.findings[2] | 0.95 | the tagger's words |
| content_tags | The demonstration cohort contained 660 stage II or III colon-cancer patients. | content_tags.findings[3] | 0.95 | the tagger's words |
| content_tags | CD3 or CD8 analysis of 21 slides and more than 2,000 cores took under four minutes per biomarker. | content_tags.findings[4] | 0.95 | the tagger's words |
| content_tags | The automated analysis counted approximately 1.2 million CD3-positive and 0.6 million CD8-positive cells. | content_tags.findings[5] | 0.95 | the tagger's words |
| content_tags | CD3 and CD8 density scores were significantly associated with disease-specific survival. | content_tags.findings[6] | 0.95 | the tagger's words |
| content_tags | Aberrant p53 immunoreactivity was associated with poorer unadjusted disease-specific survival than intermediate expression. | content_tags.findings[7] | 0.95 | the tagger's words |
| content_tags | The median PD-L1 cutoff was 1.46 percent positive cells and higher expression was associated with improved survival. | content_tags.findings[8] | 0.95 | the tagger's words |
| content_tags | PD-L1 tertiles showed a dose-response pattern and expression was primarily non-epithelial. | content_tags.findings[9] | 0.95 | the tagger's words |
| content_tags | High tumor stromal percentage was associated with lower disease-specific survival. | content_tags.findings[10] | 0.95 | the tagger's words |
| content_tags | QuPath supports fast, accurate and reproducible analysis across challenging digital-pathology applications. | content_tags.findings[11] | 0.95 | the tagger's words |
| content_tags | digital image analytics, automation, quantitation and objective tissue-sample screening also in: digital pathology | content_tags.definitions[0] | 0.95 | the tagger's words |
| content_tags | the percentage of all cells in a TMA core classified as positive also in: PD-L1 score | content_tags.definitions[2] | 0.95 | the tagger's words |
| content_tags | three times strong plus twice moderate plus once weak p53-positive tumor-nucleus percentages also in: H-score | content_tags.definitions[3] | 0.95 | the tagger's words |
| content_tags | the median of all available core scores for that patient and biomarker also in: single patient biomarker score | content_tags.definitions[4] | 0.95 | the tagger's words |
| content_tags | stromal area relative to stromal and epithelial areas also in: tumor stromal percentage | content_tags.definitions[5] | 0.95 | the tagger's words |
| content_tags | tumor epithelium, stroma and other tissue or whitespace also in: tissue areas | content_tags.definitions[7] | 0.95 | the tagger's words |
| content_tags | up to 40 GB — uncompressed data in one whole-slide z-stack plane | content_tags.quantities[0] | 0.95 | the tagger's words |
| content_tags | 660 patients — stage II and III colon-cancer demonstration cohort | content_tags.quantities[1] | 0.95 | the tagger's words |
| content_tags | 21 whole slide images — slides analyzed per CD3 or CD8 biomarker run | content_tags.quantities[2] | 0.95 | the tagger's words |
| content_tags | less than 4 minutes — runtime per biomarker for the TMA slide set | content_tags.quantities[3] | 0.95 | the tagger's words |
| content_tags | >2000 tissue cores — cores analyzed per biomarker | content_tags.quantities[4] | 0.95 | the tagger's words |
| content_tags | approximately 1.2 million — CD3-positive cells counted | content_tags.quantities[5] | 0.95 | the tagger's words |
| content_tags | 0.6 million — CD8-positive cells counted | content_tags.quantities[6] | 0.95 | the tagger's words |
| content_tags | p-values 0.006 and 0.007 — CD3 and CD8 survival-association log-rank tests | content_tags.quantities[7] | 0.95 | the tagger's words |
| content_tags | p = 0.003 — aberrant versus intermediate p53 survival comparison | content_tags.quantities[8] | 0.95 | the tagger's words |
| content_tags | 1.46% positive cells — median cutoff for the PD-L1 core score | content_tags.quantities[9] | 0.95 | the tagger's words |
| content_tags | p = 0.004 — unadjusted PD-L1 survival association | content_tags.quantities[10] | 0.95 | the tagger's words |
| content_tags | 312 patients — stage II whole-section stromal-percentage analysis | content_tags.quantities[11] | 0.95 | the tagger's words |
| content_tags | p = 0.013 — tumor-stromal-percentage survival association | content_tags.quantities[12] | 0.95 | the tagger's words |
| content_tags | 0 to 300 — H-score range from all negative to all strongly positive tumor nuclei | content_tags.quantities[13] | 0.95 | the tagger's words |
| content_tags | 40 training images — images used for interactive tumor, stroma and other classifier training | content_tags.quantities[14] | 0.95 | the tagger's words |
| content_tags | positive cell density scores also in: positive cell density scores | content_tags.variables[0] | 0.95 | the tagger's words |
| content_tags | disease-specific survival also in: disease-specific survival | content_tags.variables[1] | 0.95 | the tagger's words |
| content_tags | staining intensity also in: staining intensity | content_tags.variables[2] | 0.95 | the tagger's words |
| content_tags | maximum DAB optical density also in: maximum DAB optical density | content_tags.variables[3] | 0.95 | the tagger's words |
| content_tags | percentage of cells classified as positive | content_tags.variables[4] | 0.95 | the tagger's words |
| content_tags | tumor stromal percentage also in: tumor stromal percentage | content_tags.variables[5] | 0.95 | the tagger's words |
| content_tags | nucleus area also in: nucleus area | content_tags.variables[6] | 0.95 | the tagger's words |
| content_tags | circularity also in: circularity | content_tags.variables[7] | 0.95 | the tagger's words |
| content_tags | average hue also in: average hue | content_tags.variables[8] | 0.95 | the tagger's words |
| content_tags | Haralick texture features also in: Haralick texture features | content_tags.variables[9] | 0.95 | the tagger's words |
Tags that could not be placed in the text:
What was deliberately not tagged, and why, is recorded per paper by the tagger. Where a paper has no such record, this page says so rather than implying that nothing was declined.