1665 - QuPath versus InForm: Digital image analysis of multiplex immunofluorescence (mIF) tumor-infiltrating lymphocyte (TIL) outcomes in an immunotherapy clinical trial

Document Type


Publication Date


Publication Title

AACR Annual Meeting


oregon; portland; chiles


Background: mIF has emerged as a biomarker platform for characterizing TILs in breast cancer. One pitfall limiting wider application is the complex, proprietary nature of image analysis software. QuPath is an alternative, open-source platform recognized for its processing speed and ease of use. Here, we compare TIL analysis using QuPath version 0.2.3 versus InForm version 2.4, using specimens from a completed pre-operative phase Ib trial of locoregional cytokine therapy (IRX-2 regimen, Brooklyn ImmunoTherapeutics).
Methods: 15 paired pre-/post-immunotherapy stage I-III early stage breast cancer specimens were analyzed using the PerkinElmer Vectra platform with a 6-plex panel: cytokeratin (CK, tumor marker), CD3 (T cells), CD8 (cytotoxic T cells), CD163 (macrophages), FoxP3 (regulatory T cells, Treg), programmed death ligand 1 (PD-L1), and DAPI nuclear counterstain. Four high-power fields (HPF) were randomly selected to represent each specimen. Images were analyzed per manufacturer instructions with InForm, and corresponding data were compared with QuPath analysis. Unmixed images were imported into QuPath and cells were segmented using a confidence threshold of DAPI stain. Cells were phenotyped using basic thresholding and random trees machine learning. Tissue was segmented using machine learning. We hypothesized that QuPath and InForm would generate correlating TIL counts data and similar cohort-level estimates of immunotherapy treatment effect.
Results: We found a strong correlation in total cell count between QuPath v. InForm (R2 = 0.73, p < 0.001) and a moderate correlation in stromal/tumor area per HPF (R2 = 0.47, p < 0.001). QuPath cell phenotyping using thresholding v. random trees did not appreciably differ. However, QuPath v. InForm cell count correlations varied by cell type, with CD3+ CD8+ and CD3+ CD8- T cells strongly to moderately correlated (R2 = 0.65, p < 0.001 and R2 = 0.32, p < 0.001, respectively) and Treg and macrophage cell counts weakly correlated (R2 = 0.13, p < 0.001 and R2 = 0.08, p = 0.005). Continuous PD-L1 marker expression based on intensity was strongly correlated between softwares (μ: R2 = 0.97, p < 0.001 and σ: R2 = 0.93, p < 0.001). Estimated treatment effect on cell density varied by cell type (median fold change, QuPath v. Inform, CD8+: 3.1 v. 2.5; CD8-: 3.6 v. 1.5; Treg: 1.3 v. 1.1; macrophage: 3.2 v. 1.1). Outlier values were more common with QuPath in this analysis, resulting in unstable estimates of mean fold change.
Conclusion: Exploratory mIF biomarker outcomes in clinical trials may depend on the method of image analysis. Further work is ongoing to evaluate and improve upon discordant results using QuPath v. InForm. These data also highlight the importance of collaborative efforts via the NIH Cancer Immune Monitoring and Analysis Centers to standardize and validate mIF methodologies across institutions.

Clinical Institute



Earle A. Chiles Research Institute