Quantitative comparison of immunohistochemical staining measured by digital image analysis versus pathologist visual scoring - PubMed


Anthony E Rizzardi 1 ,

Article Images

Comparative Study

Quantitative comparison of immunohistochemical staining measured by digital image analysis versus pathologist visual scoring

Anthony E Rizzardi et al. Diagn Pathol. .

Abstract

Immunohistochemical (IHC) assays performed on formalin-fixed paraffin-embedded (FFPE) tissue sections traditionally have been semi-quantified by pathologist visual scoring of staining. IHC is useful for validating biomarkers discovered through genomics methods as large clinical repositories of FFPE specimens support the construction of tissue microarrays (TMAs) for high throughput studies. Due to the ubiquitous availability of IHC techniques in clinical laboratories, validated IHC biomarkers may be translated readily into clinical use. However, the method of pathologist semi-quantification is costly, inherently subjective, and produces ordinal rather than continuous variable data. Computer-aided analysis of digitized whole slide images may overcome these limitations. Using TMAs representing 215 ovarian serous carcinoma specimens stained for S100A1, we assessed the degree to which data obtained using computer-aided methods correlated with data obtained by pathologist visual scoring. To evaluate computer-aided image classification, IHC staining within pathologist annotated and software-classified areas of carcinoma were compared for each case. Two metrics for IHC staining were used: the percentage of carcinoma with S100A1 staining (%Pos), and the product of the staining intensity (optical density [OD] of staining) multiplied by the percentage of carcinoma with S100A1 staining (OD*%Pos). A comparison of the IHC staining data obtained from manual annotations and software-derived annotations showed strong agreement, indicating that software efficiently classifies carcinomatous areas within IHC slide images. Comparisons of IHC intensity data derived using pixel analysis software versus pathologist visual scoring demonstrated high Spearman correlations of 0.88 for %Pos (p < 0.0001) and 0.90 for OD*%Pos (p < 0.0001). This study demonstrated that computer-aided methods to classify image areas of interest (e.g., carcinomatous areas of tissue specimens) and quantify IHC staining intensity within those areas can produce highly similar data to visual evaluation by a pathologist.

Virtual slides: The virtual slide(s) for this article can be found here: http://www.diagnosticpathology.diagnomx.eu/vs/1649068103671302.

PubMed Disclaimer

Figures

Figure 1
Figure 1

Manual and automated annotations of ovarian serous carcinoma. Ovarian serous carcinoma TMA spots immunohistochemically stained for S100A1. Representative lowly and highly stained spots are shown (A-B). Image data were processed by both manual pathologist-supervised hand annotations and automated Genie Histology Pattern Recognition software. Digital hand annotations are presented as green outlines of carcinoma, excluding stroma and minimizing background and glass (C-D). These same TMA spots were classified by Genie as carcinoma (dark blue), stroma (yellow), and glass (light blue) (E-F).

Figure 2
Figure 2

Bland-Altman plots comparing automated IHC measurements (%Pos) by Hand Annotation or Genie Annotation by TMA. Bland-Altman difference plots between hand-annotated carcinomatous areas and Genie-annotated carcinomatous areas were generated for %Pos obtained using the Color Deconvolution algorithm. Data are displayed separately for TMA 1 on which the software methods were trained and TMAs 2-4 which were independent data sets. Red lines indicate mean and ± 2*standard deviation.

Figure 3
Figure 3

Bland-Altman plots comparing automated IHC measurements (OD*%Pos) by Hand Annotation or Genie Annotation by TMA. Bland-Altman difference plots between hand-annotated carcinomatous areas and Genie-annotated carcinomatous areas were generated for OD*%Pos obtained using the Color Deconvolution algorithm. Data are displayed separately for TMA 1 on which the software methods were trained and TMAs 2-4 which were independent data sets. Red lines indicate mean and ± 2*standard deviation.

Figure 4
Figure 4

Representative comparisons of pathologist visual scoring with automated IHC measurement. Ovarian serous carcinoma TMA spots stained for S100A1 were interpreted by pathologist visual scoring as 0 (no staining), 1 (<10% of carcinoma staining), 2 (10%-50% of carcinoma staining), or 3 (>50% of carcinoma staining). Representative spot for each score is shown as A-D; each column shows the identical TMA spot processed by digital methods. Genie Histology Pattern Recognition software classified tissue areas into carcinoma (dark blue), stroma (yellow), or glass (light blue) (E-H). Color Deconvolution software individually analyzed DAB staining (deconvolved by its RGB color components; I-L), and measured staining intensity only within areas classified as carcinoma. Pseudocolors represent staining intensity in shown as M-P (gray = image areas not annotated by Genie as carcinoma and therefore not considered; blue = no staining, yellow = low intensities, orange = medium intensities, and red = high intensities in Genie-annotated carcinomatous areas considered).

Figure 5
Figure 5

Automated IHC measurements (%Pos) versus pathologist visual score displayed separately for each TMA. Box plots of %Pos data generated using Genie Histology Pattern Recognition software and Color Deconvolution software within carcinomatous areas (vertical axes) versus pathologist visual score (horizontal axes). Data are displayed separately for TMA 1 on which the software methods were trained and TMAs 2-4 which were independent data sets.

Figure 6
Figure 6

Automated IHC measurements (OD*%Pos) versus pathologist visual score displayed separately for each TMA. Box plots of OD*%Pos data generated using Genie Histology Pattern Recognition software and Color Deconvolution software within carcinomatous areas (vertical axes) versus pathologist visual score (horizontal axes). Data are displayed separately for TMA 1 on which the software methods were trained and TMAs 2-4 which were independent data sets.

Similar articles

Cited by

References

    1. Schmechel SC, LeVasseur RJ, Yang KH, Koehler KM, Kussick SJ, Sabath DE. Identification of genes whose expression patterns differ in benign lymphoid tissue and follicular, mantle cell, and small lymphocytic lymphoma. Leukemia. 2004;18:841–855. doi: 10.1038/sj.leu.2403293. - DOI - PubMed
    1. Tu IP, Schaner M, Diehn M, Sikic BI, Brown PO, Botstein D, Fero MJ. A method for detecting and correcting feature misidentification on expression microarrays. BMC Genomics. 2004;5:64. doi: 10.1186/1471-2164-5-64. - DOI - PMC - PubMed
    1. Kapur K, Jiang H, Xing Y, Wong WH. Cross-hybridization modeling on Affymetrix exon arrays. Bioinformatics. 2008;24:2887–2893. doi: 10.1093/bioinformatics/btn571. - DOI - PMC - PubMed
    1. Norris AW, Kahn CR. Analysis of gene expression in pathophysiological states: balancing false discovery and false negative rates. Proc Natl Acad Sci U S A. 2006;103:649–653. doi: 10.1073/pnas.0510115103. - DOI - PMC - PubMed
    1. Freedman AN, Seminara D, Gail MH, Hartge P, Colditz GA, Ballard-Barbash R, Pfeiffer RM. Cancer risk prediction models: a workshop on development, evaluation, and application. J Natl Cancer Inst. 2005;97:715–723. doi: 10.1093/jnci/dji128. - DOI - PubMed

Publication types

MeSH terms

Substances