Abstract
Introduction
In triple-negative breast cancer (TNBC), checkpoint immunotherapy helps some patients and not others, and the protein stain used to select them, programmed death-ligand 1 (PD-L1), predicts response only loosely. Where immune cells sit may explain some of the gap: lymphocytes stalled in surrounding tissue behave differently from those touching tumor cells. Multiplex immunofluorescence can measure that, staining one section for eight proteins so every cell is typed and placed. The catch is the analysis. Cell counts rest on a chain of software decisions. Reproducibility studies show those decisions can shift the fraction of cells called positive by more than 25 percent, and they are rarely reported. We built an open pipeline that makes each decision explicit and measured what each one costs.
Methods
The pipeline runs in four stages: assemble the image, outline every nucleus with a deep learning model, measure and type each cell, then describe where the cells sit. Three choices depart from common practice. Positivity for a protein is decided per image, at the dip between that image's stained and unstained cells, rather than by a fixed cutoff or a fixed percentage. Tumor regions follow the real outline of the tumor, gaps and all, and each stretch of boundary is checked against the stain underneath before it is accepted. A cell carrying both tumor and immune signal keeps both labels and lands in a dual-identity category, infiltrated tumor or tumor-resident immune, according to which evidence is stronger. The pipeline also labels any measurement it cannot support, so unreliable numbers travel with the warning attached.
Results
We processed 279 images from 43 patients plus seven tonsil controls, typing 1,018,945 cells in one run with no failures. Each departure changed the answer. A fixed cutoff, still the common default, discarded 55 to 74 percent of visibly stained tumor cells on tumor dense images. Approximating a tumor boundary with a simple convex outline swept connective tissue immune cells inside the tumor and inflated measured infiltration nearly fourfold. Counting them all as immune mislabeled a median 76 percent of supposed infiltrating immune cells, which were tumor cells. The warning system flagged two measurements as unusable. Read from a rendered image, one protein marked 1 percent of cells positive, against 23 percent on the original unmixed signal and a published normal range of 5 to 30 percent. A second protein came from a neighboring slice never aligned to the panel, so the pipeline reports it per image only, never per cell. In the TNBC samples, immune cells were two and a half to three times denser outside tumors than inside. Suppressive T cells concentrated within the tumor mass. Neither pattern appears in tonsil.
Conclusion
The analysis steps behind a multiplex image can change a result more than the biology being compared, and most stay invisible once the numbers are written down. This tool makes them explicit and checkable. The pipeline, its tests, and the per cell measurements for all 279 images are available for reuse.