Supplementary Information

Table S1. 15 Image Analysis Features Utilized by the Risk Classifier
Biomarker / Image Analysis Feature
p53 / p53 nuclear sum intensity
p53 / p53 nuclear mean intensity
HER2/neu and Cytokeratin-20 (K20) / Ratio of mean HER2/neuintensity:mean K20 intensity in nuclei clustersb
HER2/neu and K20 / Ratio of 95th quantile HER2/neu intensity:95th quantile K20 intensity in nuclei clusters
COX2 and CD68 / Coexpression cellular COX2 mean intensity and cellular CD68 mean intensity
p53 / p53 mean intensity in nuclei clusters
p53, p16 and Hoechsta/nuclear morphology / Nuclear solidity in p53+ p16- cells
CD45RO / CD45RO plasma membrane sum intensity
AMACR / AMACR microenvironment standard deviation
COX2 / COX2 texture in cytoplasmc
HIF-1 / HIF-1 microenvironment cell mean intensity
HIF-1 / HIF-1 microenvironment cell moment (product of mean and standard deviation)
p16 / p16 cytoplasm mean intensity
p53, p16 and Hoechst/nuclear morphology / Nuclear area in p53+ p16- cells
Hoechst/nuclear morphology / Hoechst nuclear 95th quantile intensity
aHoechst is used to label DNA, which enables image analysis software to segment nuclei as individual objects within tissue images, and to measure nuclear features including area, solidity and DNA content;b Nuclei clusters are detected by the image analysis software as described by Prichard et al (1), and biomarkers are measured within the image regions containing nuclei clusters, see also Supplementary Figure S1; cContrast textural feature is extracted from a co-occurrence matrix and is a measure of the COX2 intensity contrast between a pixel and its neighbor over the whole tissue image, as described byHaralick et al(2).

Table S1

Table S2. Performance of Risk Classes Predicted by Test vs. Pathologic Diagnosis in Stratifying BE Patients with Prevalent HGD/EAC from Non-Progressor BE Patients.
A. Predictive Performance of Risk Classes vs. Generalist Pathologist Diagnosis
Variable / Hazard Ratio (95% CI) / P Value
Analysis without Risk Prediction Test
General Pathologist's Dx(LGD vs. ND/IND) / 9.65 (3.39 - 27.47) / <0.0001
Analysis with Risk Prediction Test
General Pathologist's Dx(LGD vs. ND/IND) / 3.47 (1.15 - 10.48) / 0.03
Risk Classes (predicted by the test)
Intermediate vs. Low Risk / 9.99 (1.84 - 54.2) / 0.01
High vs. Low Risk / 21.25 (4.26 - 105.86) / 0.0002
B. Predictive Performance of Risk Classes vs. GI Subspecialist Pathologist Diagnosis
Variable / Hazard Ratio (95% CI) / P Value
Analysis without Risk Prediction Test
GI Subspecialist Pathologist's Dx(LGD vs. ND/IND) / 12.95 (6.24 - 26.89) / <0.0001
Analysis with Risk Prediction Test
GI Subspecialist Pathologist's Dx(LGD vs. ND/IND) / 3.87 (1.71 - 8.77) / 0.001
Risk Classes (predicted by the test)
Intermediate vs. Low Risk / 5.37 (1.66 - 17.4) / 0.01
High vs. Low Risk / 12.46 (4.11 - 37.76) / <0.0001
Multivariate Cox models were run in which subsequent diagnosis of HGD/EAC was evaluated first in relation to pathologic diagnosis alone, then in relation to risk classes and pathologic diagnosis in non-progressorsand prevalent cases. Variables were dichotomized as follows; diagnosis: LGD vs. ND and IND combined, risk classes predicted by the test: intermediate vs. low risk class and high vs. low risk class. Part A, n=130 patients with generalist diagnosis recorded during surveillance. Part B, n=175 patients with GI subspecialist diagnosis provided for this study (i.e. all patients).

Figure S1

Figure S1. Representative Whole Slide Scan and Image Analysis Masks. A: Whole slide image of a BE biopsy specimen fluorescently immunolabeled for HER2/neu (green), K20 (red) and Hoechst (blue), which is provided as an example panel of biomarkers; B: 20x zoom view of theslide image region highlighted by the purple box; C:Image objects segmented as nuclei by the image analysis software based on the Hoechst channel; D: Objects segmented as whole cell by the software; E: Objects segmented as cytoplasm by subtracting nuclei area from whole cell area; F: Objects segmented as plasma membrane based on the HER2 channel; G: Objects segmented as nuclei clusters based on the Hoechst channel and morphology. Once nuclei, cell, cytoplasm, plasma membrane and nuclei cluster objects are segmented/detected, the image analysis software calculates features on biomarkers and morphology, e.g. mean biomarker intensity in nuclei clusters, mean biomarker intensity in nuclei, sum intensity of biomarker in plasma membrane, mean biomarker intensity in cytoplasm, etc (see list in Table S1 and detailed descriptions of the image analysis algorithms in (1)).

References cited in Supplementary Material

1.Prichard, J. W., Davison, J. M., Campbell, B. B., Repa, K. A., Reese, L. M., Nguyen, X. M., Li, J., Foxwell, T., Taylor, D. L., and Critchley-Thorne, R. J. TissueCypher: A Systems Biology Approach to Anatomic Pathology. Journal of Pathology Informatics, 6:48., 2015.

2.Haralick, R. M., Shanmugam, K., and Dinstein, I. Textural Feature for Image Classification. IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, SMC-3: 610-621, 1973.